{ "paper_id": "J03-3003", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T02:15:33.953963Z" }, "title": "Embedding Web-Based Statistical Translation Models in Cross-Language Information Retrieval", "authors": [ { "first": "Wessel", "middle": [], "last": "Kraaij", "suffix": "", "affiliation": { "laboratory": "", "institution": "TNO TPD", "location": { "postBox": "PO BOX 155", "postCode": "2600 AD", "settlement": "Delft", "country": "The Netherlands" } }, "email": "kraaij@tpd.tno.nl" }, { "first": "Jian-Yun", "middle": [], "last": "Nie", "suffix": "", "affiliation": {}, "email": "nie@iro.umontreal.ca" }, { "first": "Michel", "middle": [], "last": "Simard", "suffix": "", "affiliation": {}, "email": "simardm@iro.umontreal.ca" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Although more and more language pairs are covered by machine translation (MT) services, there are still many pairs that lack translation resources. Cross-language information retrieval (CLIR) is an application that needs translation functionality of a relatively low level of sophistication, since current models for information retrieval (IR) are still based on a bag of words. The Web provides a vast resource for the automatic construction of parallel corpora that can be used to train statistical translation models automatically. The resulting translation models can be embedded in several ways in a retrieval model. In this article, we will investigate the problem of automatically mining parallel texts from the Web and different ways of integrating the translation models within the retrieval process. Our experiments on standard test collections for CLIR show that the Web-based translation models can surpass commercial MT systems in CLIR tasks. These results open the perspective of constructing a fully automatic query translation device for CLIR at a very low cost.", "pdf_parse": { "paper_id": "J03-3003", "_pdf_hash": "", "abstract": [ { "text": "Although more and more language pairs are covered by machine translation (MT) services, there are still many pairs that lack translation resources. Cross-language information retrieval (CLIR) is an application that needs translation functionality of a relatively low level of sophistication, since current models for information retrieval (IR) are still based on a bag of words. The Web provides a vast resource for the automatic construction of parallel corpora that can be used to train statistical translation models automatically. The resulting translation models can be embedded in several ways in a retrieval model. In this article, we will investigate the problem of automatically mining parallel texts from the Web and different ways of integrating the translation models within the retrieval process. Our experiments on standard test collections for CLIR show that the Web-based translation models can surpass commercial MT systems in CLIR tasks. These results open the perspective of constructing a fully automatic query translation device for CLIR at a very low cost.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Finding relevant information in any language on the increasingly multilingual World Wide Web poses a real challenge for current information retrieval (IR) systems. We will argue that the Web itself can be used as a translation resource in order to build effective cross-language IR systems.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1." }, { "text": "The goal of IR is to find relevant documents from a large collection of documents or from the World Wide Web. To do this, the user typically formulates a query, often in free text, to describe the information need. The IR system then compares the query with each document in order to evaluate its similarity (or probability of relevance) to the query. The retrieval result is a list of documents presented in decreasing order of similarity. The key problem in IR is that of effectiveness, that is, how good an IR system is at retrieving relevant documents and discarding irrelevant ones.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Information Retrieval and Cross-Language Information Retrieval", "sec_num": "1.1" }, { "text": "Because of the information explosion that has occurred on the Web, people are more in need of effective IR systems than ever before. The search engines currently available on the Web are IR systems that have been created to answer this need. By querying these search engines, users are able to identify quickly documents containing the same keywords as the query they enter. However, the existing search engines provide only monolingual IR; that is, they retrieve documents only in the same lan-guage as the query. To be more precise: Search engines usually do not consider the language of the keywords when the keywords of a query are matched against those of the documents. Identical keywords are matched, whatever their languages are. For example, the English word son can match the French word son ('his' or 'her') . Current search engines do not provide the functionality for cross-language IR (CLIR), that is, the ability to retrieve relevant documents written in languages different from that of the query (without the query's being translated manually into the other language(s) of interest).", "cite_spans": [ { "start": 802, "end": 818, "text": "('his' or 'her')", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Information Retrieval and Cross-Language Information Retrieval", "sec_num": "1.1" }, { "text": "As the Web has grown, more and more documents on the Web have been written in languages other than English, and many Internet users are non-native English speakers. For many users, the barrier between tbe language of the searcher and the langage in which documents are written represents a serious problem. Although many users can read and understand rudimentary English, they feel uncomfortable formulating queries in English, either because of their limited vocabulary in English, or because of the possible misusage of English words. For example, a Chinese user may use economic instead of cheap or economical or inexpensive in a query because these words have a similar translation in Chinese. An automatic query translation tool would be very helpful to such a user. On the other hand, even if a user masters several languages, it is still a burden for him or her to formulate several queries in different languages. A query translation tool would also allow such a user to retrieve relevant documents in all the languages of interest with only one query. Even for users with no understanding of a foreign language, a CLIR system might still be useful. For example, someone monitoring a competitor's developments with regard to products similar to those he himself produces might be interested in retrieving documents describing the possible products, even if he does not understand them. Such a user might use machine translation systems to get the gist of the contents of the documents he retrieves through his query. For all these types of users, CLIR would represent a useful tool.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Information Retrieval and Cross-Language Information Retrieval", "sec_num": "1.1" }, { "text": "From an implementation point of view, the only difference between CLIR and the classical IR task is that the query language differs from the document language. It is obvious that to perform in an effective way the task of retrieving documents that are relevant to a query when the documents are written in a different language than the query, some form of translation is required. One might conjecture that a combination of two existing fields, IR and machine translation (MT), would be satisfactory for accomplishing the combined translation and retrieval task. One could simply translate the query by means of an MT system, then use existing IR tools, obviating the need for a special CLIR system. This approach, although feasible, is not the only possible approach, nor is it necessarily the best one. MT systems try to translate text into a well-readable form governed by morphological, syntactic, and semantic constraints. However, current IR models are based on bag-of-words models. They are insensitive to word order and to the syntactic structure of queries. For example, with current IR models, the query \"computer science\" will usually produce the same retrieval results as \"science computer.\" The complex process used in MT for producing a grammatical translation is not fully exploited by current IR models. This means that a simpler translation approach may suffice to implement the translation step.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Possible Approaches to CLIR", "sec_num": "1.2" }, { "text": "On the other hand, MT systems are far from perfect. They often produce incorrect translations. For example, Systran 1 translates the word drug as drogue (illegal substance) in French for both drug traffic and drug administration office. Such a translation error will have a substantial negative impact on the effectiveness of any CLIR system that incorporates it. So even if MT systems are used as translation devices, they may need to be complemented by other, more robust translation tools to improve their effectiveness.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Possible Approaches to CLIR", "sec_num": "1.2" }, { "text": "In the current study, we will use statistical translation models as such a complementary tool. Queries submitted to IR systems or search engines are often very short. In particular, the average length of queries submitted to the search engines on the Web is about two words (Jansen et al. 2001) . Such short queries are generally insufficient to describe the user's information need in a precise and unambiguous way. Many important words are missing from them. For example, a user might formulate the query \"Internet connection\" in order to retrieve documents about computer networks, Internet service providers, or proxies. However, under the current bag-of-words approach, the relevant documents containing these terms are unlikely to be retrieved. To solve this problem, a common approach used in IR is query expansion, which tries to add synonyms or related words to the original query, making the expanded query a more exhaustive description of the information need. The words added to the query during query expansion do not need to be strict synonyms to improve the query results. However, they do have to be related, to some degree, to the user's information need. Ideally, the degree of the relatedness should be weighted, with a strongly related word weighted more heavily in the expanded query than a less related one.", "cite_spans": [ { "start": 274, "end": 294, "text": "(Jansen et al. 2001)", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Possible Approaches to CLIR", "sec_num": "1.2" }, { "text": "MT systems act in a way opposite to the query expansion process: Only one translation is generally selected to express a particular meaning. 2 In doing so, MT systems employed in IR systems in fact restrict the possible query expansion effect during the translation process. We believe that CLIR can benefit from query translation that provides multiple translations for the same meaning. In this regard, the tests carried out by Kwok (1999) with a commercial MT system for Chinese-English CLIR are quite interesting. His experiments show that it is much better to use the intermediate translation data produced by the MT system than the final translation itself. The intermediate data contain, among other things, all the possible translation words for query terms. Kwok's work clearly demonstrates that using an MT system as a black box is not the most effective choice for query translation in CLIR. However, few MT systems allow one to access the intermediate stages of the translations they produce.", "cite_spans": [ { "start": 141, "end": 142, "text": "2", "ref_id": null }, { "start": 430, "end": 441, "text": "Kwok (1999)", "ref_id": "BIBREF20" } ], "ref_spans": [], "eq_spans": [], "section": "Possible Approaches to CLIR", "sec_num": "1.2" }, { "text": "Apart from the MT approach, queries can also be translated by using a machinereadable bilingual dictionary or by exploiting a set of parallel texts (texts with their translations). High-quality bilingual dictionaries are expensive, but there are many free on-line translation dictionaries available on the Web that can be used for query translation. This approach has been applied in several studies (e.g., Hull and Grefenstette 1996; Hiemstra and Kraaij 1999) . However, free bilingual dictionaries often suffer from a poor coverage of the vocabulary in the two languages with which they deal, and from the problem of translation ambiguity, because usually no information is provided to allow for disambiguation. Several previous studies (e.g., Nie et al. 1999) , have shown that using a translation dictionary alone would produce much lower effectiveness than an MT system. However, a dictionary complemented by a statistical language model (Gao et al. 2001; Xu, Weischedel, and Nguyen 2001) has produced much better results than when the dictionary is used alone.", "cite_spans": [ { "start": 407, "end": 434, "text": "Hull and Grefenstette 1996;", "ref_id": "BIBREF10" }, { "start": 435, "end": 460, "text": "Hiemstra and Kraaij 1999)", "ref_id": null }, { "start": 746, "end": 762, "text": "Nie et al. 1999)", "ref_id": "BIBREF35" }, { "start": 943, "end": 960, "text": "(Gao et al. 2001;", "ref_id": null }, { "start": 961, "end": 993, "text": "Xu, Weischedel, and Nguyen 2001)", "ref_id": "BIBREF50" } ], "ref_spans": [], "eq_spans": [], "section": "Possible Approaches to CLIR", "sec_num": "1.2" }, { "text": "In this article, the use of a bilingual dictionary is not our focus. We will concentrate on a third alternative for query translation: an approach based on parallel texts. Parallel texts are texts accompanied by their translation in one or several other languages (V\u00e9ronis 2000) . They contain valuable translation examples for both human and machine translation. A number of studies in recent years (e.g., Nie et al. 1999; Franz et al. 2001; Sheridan, Ballerini, and Sch\u00e4uble 1998; Yang et al. 1998) have explored the possibility of using parallel texts for query translation in CLIR. One potential advantage of such an approach is that it provides multiple translations for the same meaning. The translation of a query would then contain not only words that are true translations of the query, but also related words. This is the query expansion effect that we want to produce in IR. Our experimental results have confirmed that this approach can be very competitive with the MT approach and yield much better results than a simple dictionary-based approach, while keeping the development cost low.", "cite_spans": [ { "start": 264, "end": 278, "text": "(V\u00e9ronis 2000)", "ref_id": null }, { "start": 407, "end": 423, "text": "Nie et al. 1999;", "ref_id": "BIBREF35" }, { "start": 424, "end": 442, "text": "Franz et al. 2001;", "ref_id": null }, { "start": 443, "end": 482, "text": "Sheridan, Ballerini, and Sch\u00e4uble 1998;", "ref_id": "BIBREF44" }, { "start": 483, "end": 500, "text": "Yang et al. 1998)", "ref_id": "BIBREF51" } ], "ref_spans": [], "eq_spans": [], "section": "Possible Approaches to CLIR", "sec_num": "1.2" }, { "text": "However, one major obstacle to the use of parallel texts for CLIR is the unavailability of large parallel corpora for many language pairs. Hence, our first goal in the research presented here was to develop an automatic mining system that collects parallel pages on the Web. The collected parallel Web pages are used to train statistical translation models (TMs) that are then applied to query translation. Such an approach offers the advantage of enabling us to build a CLIR system for a new language pair without waiting for the release of an MT system for that language pair. The number of potential language pairs supported by Web-based translation models is large if one includes transitive translation using English as a pivot language. English is often one of the languages of those Web pages for which parallel translations are available.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Possible Approaches to CLIR", "sec_num": "1.2" }, { "text": "The main objectives of this article are twofold: (1) We will show that it is possible to obtain large parallel corpora from the Web automatically that can form the basis for an effective CLIR system, and (2) we will compare several ways to embed translation models in an IR system to exploit these corpora for cross-language query expansion.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Possible Approaches to CLIR", "sec_num": "1.2" }, { "text": "Our experiments will show that these translation tools can result in CLIR of comparable effectiveness to MT systems. This in turn will demonstrate the feasibility of exploiting the Web as a large parallel corpus for the purpose of CLIR.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Possible Approaches to CLIR", "sec_num": "1.2" }, { "text": "Now let us turn to query translation problems. Previous studies on CLIR have identified three problems for query translation (Grefenstette 1998): identifying possible translations, pruning unlikely translations, and weighting the translation words.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problems in Query Translation", "sec_num": "1.3" }, { "text": "Finding translations. First of all, whatever translation tool is employed in translating queries has to provide a good coverage of the source and target vocabularies. In a dictionary-based approach to CLIR, we will encounter the same problems that have been faced in MT research: phrases, collocations, idioms, and domain-specific terminology are often translated incorrectly. These classes of expressions require a sophisticated morphological analysis, and furthermore, domain-specific terms challenge the lexical coverage of the transfer dictionaries. A second important class of words that can pose problems for CLIR, particularly that involving news article retrieval, is proper names. The names of entities such as persons or locations are frequently used in queries for news articles, and their translation is not always trivial. Often, the more commonly used geographical names of countries or their capitals have a different spelling in different languages (e.g., Milan/Milano/Milaan) or translations that are not related to the same morphological root (e.g., Germany/Allemagne/Duitsland). The names of organizations and their abbreviations are also a notorious problem; for example, the United Nations can be referred to as UN, ONU, VN, etc. (disregarding the problem of morphological normalization of abbreviations). When proper names have to be translated from one language to another with a different script, like Cyrillic, Arabic, or Chinese, this problem is even more acute. The process of defining the spelling of a named entity in a language with a different script from the originating language is called transliteration and is based on a phonemic representation of the named entity. Unfortunately different national \"standards\" are used for transliteration in different languages that use the same alphabet (e.g., the former Russian president's name in Latin script has been transliterated as Jeltsin, Eltsine, Yeltsin, and Jelzin.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problems in Query Translation", "sec_num": "1.3" }, { "text": "Pruning translation alternatives. A word or a term often has multiple translations. Some of them are appropriate for a particular query and the others are not. An important question is how to keep the appropriate translations while eliminating the inappropriate ones. Because of the particularities of IR, it might improve the results to retain multiple translations that display small differences in sense, as in query expansion. So it could be beneficial to keep all related senses for the matching process, together with their probabilities.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problems in Query Translation", "sec_num": "1.3" }, { "text": "Weighting translation alternatives. Closely related to the previous point is the question of how to deal with translation alternatives. The weighting of words in documents and in the query is of crucial importance in IR. A word with a heavy weight will influence the results of retrieval more than a low-weight word. In CLIR it is also important to assign appropriate weights to translation words. Pruning translations can be viewed as an extreme Boolean way of weighting translations. The intuition is that, just as in query expansion, it may well be beneficial to assign a heavier weight to the \"main\" translation and a lighter weight to related translations.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problems in Query Translation", "sec_num": "1.3" }, { "text": "The problem of \"weighting of translation alternatives,\" identified by Grefenstette, refers to the more general problem of designing an architecture for a CLIR system in which translation and document ranking are integrated in a way that maximizes retrieval effectiveness.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Integration of Query Translation with Retrieval", "sec_num": "1.4" }, { "text": "The MT approach clearly separates translation from retrieval: A query is first translated, and the result of the translation is subsequently submitted to an IR system as a new query. At the retrieval phase, one no longer knows how certain a translated word is with respect to the other translated words in the translated query. All the translation words are treated as though they are totally certain. Indeed, an MT system is used as a black box. In this article, we consider translation to be an integral part of the IR process that has to be considered together with the retrieval step.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Integration of Query Translation with Retrieval", "sec_num": "1.4" }, { "text": "From a more theoretical point of view, CLIR is a process that, taken as a whole, is composed of query translation, document indexing, and document matching. The two first subprocesses try to transform the query and the documents into a comparable internal representation form. The third subprocess tries to compare the representations to evaluate the similarity. In previous studies on CLIR, the first subprocess is clearly separated from the latter two, which are integrated in classical IR systems. An approach that considers all three subprocesses together will have the advantage of accounting better for the uncertainty of translation during retrieval. More analysis on this point is provided in Nie (2002) . This article follows the same direction as Nie's. We will show in our experiments that an integrated approach can produce very high CLIR effectiveness.", "cite_spans": [ { "start": 701, "end": 711, "text": "Nie (2002)", "ref_id": "BIBREF32" } ], "ref_spans": [], "eq_spans": [], "section": "Integration of Query Translation with Retrieval", "sec_num": "1.4" }, { "text": "An attractive framework for integrating translation and retrieval is the probabilistic framework, although estimating translation probabilities is not always straightforward using this framework.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Integration of Query Translation with Retrieval", "sec_num": "1.4" }, { "text": "In summary, because CLIR does not necessarily require a unique translation of a text (as MT does), approaches other than fully automatic MT might provide interesting characteristics for CLIR that are complementary to those of MT approaches. This could result in greater precision, 3 since an MT system might choose the wrong translation for the query term(s), and/or a higher rate of recall, 4 since multiple translations are accommodated, which could retrieve documents via related terminology.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Integration of Query Translation with Retrieval", "sec_num": "1.4" }, { "text": "In this article we will investigate the effectiveness of CLIR systems based on probabilistic translation models trained on parallel texts mined from the Web. Globally, our approach to the CLIR problem can be viewed informally as \"cross-lingual sense matching.\" Both query and documents are modeled as a distribution over semantic concepts, which in reality is approximated by a distribution over words. The challenge for CLIR is to measure to what extent these concepts (or word senses) are related. From this point of view, our approach is similar in principle to that using latent semantic analysis (LSI) (Dumais et al. 1997) , which also tries to create semantic similarity between documents, queries, and terms by transposing them into a new vector space. An alternative way of integrating translation and IR is to create \"structured queries,\" in which translations are modeled as synonyms (Pirkola 1998) . Since this approach is simple and effective, we will use it as one of the reference systems in our experiments.", "cite_spans": [ { "start": 607, "end": 627, "text": "(Dumais et al. 1997)", "ref_id": null }, { "start": 894, "end": 908, "text": "(Pirkola 1998)", "ref_id": "BIBREF36" } ], "ref_spans": [], "eq_spans": [], "section": "Integration of Query Translation with Retrieval", "sec_num": "1.4" }, { "text": "The general approach of this article will be implemented in several different ways, each fully embedded in the retrieval models tested. A series of experiments on CLIR will be conducted in order to evaluate these models. The results will clearly show that Web-based translation models are as effective as (and sometimes more effective than) off-the-shelf commercial MT systems.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Integration of Query Translation with Retrieval", "sec_num": "1.4" }, { "text": "The remainder of the article is organized as follows: Section 2 discusses the procedure we used to construct parallel corpora from the Web, and Section 3 describes the procedure we used to train the translation models. Section 4 describes the probabilistic IR model that we employed and various ways of embedding translation into a retrieval model. Section 5 presents our experimental results. The article ends with a discussion and conclusion section.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Integration of Query Translation with Retrieval", "sec_num": "1.4" }, { "text": "It has been shown that by using a large parallel corpus, one can produce CLIR effectiveness close to that obtained with an MT system (Nie et al. 1999) . In previous studies, parallel texts have been exploited in several ways: using a pseudofeedback approach, capturing global cross-language term associations, transposing to a languageindependent semantic space, and training a statistical translation model. Using a pseudofeedback approach. In Yang et al. (1998) parallel texts are used as follows. A given query in the source language is first used to retrieve a subset of texts from the parallel corpus. The corresponding subset in the target language is considered to provide a description of the query in the target language. From this subset of documents, a set of weighted words is extracted, and this set of words is used as the query \"translation.\"", "cite_spans": [ { "start": 133, "end": 150, "text": "(Nie et al. 1999)", "ref_id": "BIBREF35" }, { "start": 445, "end": 463, "text": "Yang et al. (1998)", "ref_id": "BIBREF51" } ], "ref_spans": [], "eq_spans": [], "section": "PTMiner", "sec_num": "2." }, { "text": "Capturing global cross-language term associations. A more advanced and theoretically better-motivated approach is to index concatenated parallel documents in the dual space of the generalized vector space model (GVSM), where terms are indexed by documents (Yang et al. 1998 ). An approach related to GVSM is to build a so-called similarity thesaurus on the parallel or comparable corpus. A similarity thesaurus is an information structure (also based on the dual space of indexing terms by documents) in which associated terms are computed on the basis of global associations between terms as measured by term co-occurrence on the document level (Sheridan, Ballerini, and Sch\u00e4uble 1998) . Recently, the idea of using the dual space of parallel documents for cross-lingual query expansion has been recast in a language-modeling framework (Lavrenko, Choquette, and Croft 2002) .", "cite_spans": [ { "start": 256, "end": 273, "text": "(Yang et al. 1998", "ref_id": "BIBREF51" }, { "start": 646, "end": 686, "text": "(Sheridan, Ballerini, and Sch\u00e4uble 1998)", "ref_id": "BIBREF44" }, { "start": 837, "end": 874, "text": "(Lavrenko, Choquette, and Croft 2002)", "ref_id": "BIBREF24" } ], "ref_spans": [], "eq_spans": [], "section": "PTMiner", "sec_num": "2." }, { "text": "Transposing to a language-independent semantic space. The concatenated documents can also be transposed in a language-independent space by applying latent semantic indexing (Dumais et al. 1997; Yang et al. 1998) . The disadvantage of this approach is that the concepts in this space are hard to interpret and that LSI is computationally demanding. It is currently not feasible to perform such a transposition on a Web scale.", "cite_spans": [ { "start": 173, "end": 193, "text": "(Dumais et al. 1997;", "ref_id": null }, { "start": 194, "end": 211, "text": "Yang et al. 1998)", "ref_id": "BIBREF51" } ], "ref_spans": [], "eq_spans": [], "section": "PTMiner", "sec_num": "2." }, { "text": "Training a statistical translation model. Approaches that involve training a statistical translation model have been explored in, for example, Nie et al. (1999) and Franz et al. (2001) . In Nie et al.'s approach, statistical translation models (usually IBM model 1) are trained on a parallel corpus. The models are used in a straightforward way: The source query is submitted to the translation model, which proposes a set of translation equivalents, together with their probability. The latter are then used as a query for the retrieval process, which is based on a vector space model. Franz et al.'s approach uses a better founded theoretical framework: the OKAPI probabilistic IR model (Robertson and Walker 1994) . The present study uses a different probabilistic IR model, one based on statistical language models (Hiemstra 2001; Xu, Weischedel, and Nguyen 2001 ). This IR model facilitates a tighter integration of translation and retrieval. An important difference between statistical translation approaches and approaches based on document alignment discussed in the previous paragraph is that translation models perform alignment at a much more refined level. Consequently, the alignments can be used to estimate translation relations in a reliable way. On the other hand, the advantage of the CLIR approaches that rely simply on alignment at the document level is that they can also handle comparable corpora, that is, documents that discuss the same topic but are not necessarily translations of each other (Laffling 1992) .", "cite_spans": [ { "start": 143, "end": 160, "text": "Nie et al. (1999)", "ref_id": "BIBREF35" }, { "start": 165, "end": 184, "text": "Franz et al. (2001)", "ref_id": null }, { "start": 689, "end": 716, "text": "(Robertson and Walker 1994)", "ref_id": "BIBREF41" }, { "start": 819, "end": 834, "text": "(Hiemstra 2001;", "ref_id": null }, { "start": 835, "end": 866, "text": "Xu, Weischedel, and Nguyen 2001", "ref_id": "BIBREF50" }, { "start": 1518, "end": 1533, "text": "(Laffling 1992)", "ref_id": "BIBREF23" } ], "ref_spans": [], "eq_spans": [], "section": "PTMiner", "sec_num": "2." }, { "text": "Most previous work on parallel texts has been conducted on a few manually constructed parallel corpora, notably the Canadian Hansard corpus. This corpus 5 contains many years' debates in the Canadian parliament in both English and French, amounting to several dozens of millions of words in each language. The European parliament documents represent another large parallel corpus in several European languages. However, the availability of this corpus is much more restricted than the Canadian Hansard. The Hong Kong government publishes official documents in both Chinese and English. They form a Chinese-English parallel corpus, but again, its size is much smaller than that of the Canadian Hansard. For many other languages, no large parallel corpora are available for the training of statistical models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "PTMiner", "sec_num": "2." }, { "text": "LDC has tried to collect additional parallel corpora, resorting at times to manual collection (Ma 1999) . Several other research groups (for example, the RALI lab at Universit\u00e9 de Montr\u00e9al) have also tried to acquire manually constructed parallel corpora. However, manual collection of large corpora is a tedious task that is timeand resource-consuming. On the other hand, we observe that the increasing usage of different languages on the Web results in more and more bilingual and multilingual sites. Many Web pages are now translated into different languages. The Web contains a large number of parallel Web pages in many languages (usually with English). If these can be extracted automatically, then this would help solve, to some extent, the problem of parallel corpora. PTMiner (for Parallel Text Miner) was built precisely for this purpose.", "cite_spans": [ { "start": 94, "end": 103, "text": "(Ma 1999)", "ref_id": "BIBREF27" } ], "ref_spans": [], "eq_spans": [], "section": "PTMiner", "sec_num": "2." }, { "text": "Of course, an automatic mining program is unable to understand the texts it extracts and hence to judge in a totally reliable way whether they are parallel. However, CLIR is quite error-tolerant. As we will show, a noisy parallel corpus can still be very useful for CLIR.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "PTMiner", "sec_num": "2." }, { "text": "Parallel Web pages usually are not published in isolation; they are often linked to one another in some way. For example, Resnik (1998) observed that some parallel Web pages are often referenced in the same parent index Web page. In addition, the anchor text of such links usually identifies the language. For example, if a Web page index.html provides links to both English and French versions of a page it references, and the anchor texts of the links are respectively \"English version\" and \"French version,\" then the referenced versions are probably parallel pages in English and French. To locate such pages, Resnik first sends a query of the following form to the Web search engine AltaVista, which returns the parent indexing pages:", "cite_spans": [ { "start": 122, "end": 135, "text": "Resnik (1998)", "ref_id": "BIBREF40" } ], "ref_spans": [], "eq_spans": [], "section": "General Principles of Automatic Mining", "sec_num": "2.1" }, { "text": "anchor: English AND anchor: French Then the referenced pages in both languages are retrieved and considered to be parallel. Applying this method, Resnik was able to mine 2,491 pairs of English-French Web pages. Other researchers have adapted his system to mine 3,376 pairs of English-Chinese pages and 59 pairs of English-Basque pages.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "General Principles of Automatic Mining", "sec_num": "2.1" }, { "text": "We observe, however, that only a small portion of parallel Web sites are organized in this way. Many other parallel pages cannot be found with Resnik's method. The mining system we employ in the research presented here uses different criteria from Resnik's; and we also incorporate an exploration process (i.e., a host crawler) in order to discover Web pages that have not been indexed by the existing search engines.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "General Principles of Automatic Mining", "sec_num": "2.1" }, { "text": "The mining process in PTMiner is divided into two main steps: identification of candidate parallel pages, and verification of their parallelism. The overall process is organized into the following steps:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "General Principles of Automatic Mining", "sec_num": "2.1" }, { "text": "Determining candidate sites. Identify Web sites that may contain parallel pages. In our approach, we adopt a simple definition of Web site: a host corresponding to a distinct DNS (domain name system) address (e.g., www.altavista.com and geocities.yahoo.com ).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "1.", "sec_num": null }, { "text": "File name fetching. Identify a set of Web pages on each Web site that are indexed by search engines.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "2.", "sec_num": null }, { "text": "Host crawling. Use the URLs collected in the previous step as seeds to further crawl each candidate site for more URLs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "3.", "sec_num": null }, { "text": "Pair scanning by names. Construct pairs of Web pages on the basis of pattern matching between URLs (e.g., index.html vs. index f.html ).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4.", "sec_num": null }, { "text": "Text filtering. Filter the candidate parallel pages further according to several criteria that operate on their contents.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "5.", "sec_num": null }, { "text": "In the following subsections, we describe each of these steps in more detail.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "5.", "sec_num": null }, { "text": "In addition to the organization of parallel Web pages exploited by Resnik's method, another common characteristic of parallel Web pages is that they cross-reference one another. For example, an English Web page may contain a pointer to the French version, and vice versa, and the anchor text of these pointers usually indicates the language of the other page. This phenomenon is common because such an anchor text shows the reader that a version in another language is available.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Identification of Candidate Web Sites", "sec_num": "2.2" }, { "text": "In considering both ways of organizing parallel Web pages, we see that a common feature is the existence of a link with an anchor text identifying a language. This is the criterion we use in PTMiner to detect candidate Web sites: the existence of at least one Web page containing such a link. Candidate Web sites are identified via requests sent to a search engine (e.g., AltaVista or Google). For example, the following request asks for pages in English that contain a link with one of the required anchor texts.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Identification of Candidate Web Sites", "sec_num": "2.2" }, { "text": "anchor: French version, in French, en Fran\u00e7ais, . . .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Identification of Candidate Web Sites", "sec_num": "2.2" }, { "text": "The hosts extracted from the responses are considered to be candidate sites.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "language: English", "sec_num": null }, { "text": "It is assumed that parallel pages are stored on the same Web site. This is not always true, but this assumption allows us to minimize the exploration of the Web and to avoid considering many unlikely candidates.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "File Name Fetching", "sec_num": "2.3" }, { "text": "To search for parallel pairs of pages from each candidate site, PTMiner first asks the search engine for all the Web pages from a particular site that it has indexed, via a request of the following form:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "File Name Fetching", "sec_num": "2.3" }, { "text": "host: However, the results of this step may not be exhaustive, because", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "File Name Fetching", "sec_num": "2.3" }, { "text": "\u2022 search engines typically do not index all the Web pages of a site.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "File Name Fetching", "sec_num": "2.3" }, { "text": "\u2022 most search engines allow users to retrieve a limited number of documents (e.g., 1,000 in AltaVista).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "File Name Fetching", "sec_num": "2.3" }, { "text": "Therefore, we continue our search with a host crawler, which uses the Web pages found by the search engines as seeds.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "File Name Fetching", "sec_num": "2.3" }, { "text": "A host crawler is slightly different from a Web crawler or a robot in that a host crawler can only exploit one Web site at a time. A breadth-first crawling algorithm is used in the host-crawling step of PTMiner's mining process. The principle is that if a retrieved Web page contains a link to an unexplored document on the same site, this document is added to the list of pages to be explored later. This crawling step allows us to obtain more Web pages from the candidate sites.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Host Crawling", "sec_num": "2.4" }, { "text": "Once a large set of URLs has been identified, the next task is to find parallel pairs among them. In our experience, many parallel Web pages have very similar file names.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pair Scanning by Names", "sec_num": "2.5" }, { "text": "For example, an English Web page with the file name index.html often corresponds to a French translation with a file name such as index f.html . The only difference between the two file names is a segment that identifies the language of the file. This similarity in file names is by no means an accident. In fact, this is a common way for Webmasters to keep track of a large number of documents in different versions. This same observation also applies to URL paths. For example, the following two URLs are also similar in name: http : //www.asite.ca/en/afile.html and http : //www.asite.ca/fr/afile.html .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pair Scanning by Names", "sec_num": "2.5" }, { "text": "To find similarly named URLs, we define lists of prefixes and suffixes for both the source and the target languages. For example:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pair Scanning by Names", "sec_num": "2.5" }, { "text": "EnglishPrefix = {(emptychar), e, en, english, e , en , english , . . .}", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pair Scanning by Names", "sec_num": "2.5" }, { "text": "Once a possible source language prefix is identified in an URL, it is replaced with a prefix in the target language, and we then test if this URL is found on the Web site.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pair Scanning by Names", "sec_num": "2.5" }, { "text": "The file pairs identified in previous steps are further verified in regard to their contents. In PTMiner, the following criteria are used for verification: file length, HTML structure, and language and character set.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Filtering by Contents", "sec_num": "2.6" }, { "text": "The ratio of the lengths of a pair of parallel pages is usually comparable to the typical length ratio of the two languages (especially when the text is long enough). Hence, a simple verification is to compare the lengths of the two files. As many Web documents are quite short, we tolerate some difference (up to 40% from the typical ratio).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "File Length.", "sec_num": "2.6.1" }, { "text": "Parallel Web pages are usually designed to have similar layouts. This often means that the two parallel pages have similar HTML structures. However, the HTML structures of parallel pages may also be quite different from one another. Pages may look similar and still have different HTML markups. Therefore, a certain amount of flexibility is also employed in this step.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "HTML Structure.", "sec_num": "2.6.2" }, { "text": "In our approach, we first determine a set of meaningful HTML tags that affect the appearance of the page and extract them from both files (e.g.,

and

, but not and ). A \"diff\"-style comparison will reveal how different the two extracted sequences of tags are. A threshold is set to filter out the pairs of pages that are not similar enough in HTML structure.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "HTML Structure.", "sec_num": "2.6.2" }, { "text": "At this stage, nontextual parts of the pages are also removed. If a page does not contain enough text, it is also discarded.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "HTML Structure.", "sec_num": "2.6.2" }, { "text": "When we query search engines for documents in one specific language, the returned documents may actually be in a different language from the one we specified. This problem is particularly serious for Asian languages. When we ask for Chinese Web pages, we often obtain Korean Web pages, because the language of the documents has not been identified accurately by the search engines. Another, more important factor that makes it necessary to use a language detector is that during host crawling and pair scanning, no verification is done with regard to languages. All files with an en suffix in their names, for example, are assumed to be English pages, which may be an erroneous assumption.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Language and Character Set.", "sec_num": "2.6.3" }, { "text": "To filter out the files not in the required languages, the SILC system (Isabelle, Simard, and Plamondon 1998) is used. SILC employs n-gram statistical language models to determine the most probable language and encoding schema for a text. It has been trained on a number of large corpora for several languages. The accuracy of the system is very high. When a text contains at least 50 characters, its accuracy is almost perfect. SILC can filter out a set of file pairs that are not in the required languages.", "cite_spans": [ { "start": 71, "end": 109, "text": "(Isabelle, Simard, and Plamondon 1998)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Language and Character Set.", "sec_num": "2.6.3" }, { "text": "Our utilization of HTML structure to determine whether two pages are parallel is similar to that of Resnik (1998) , who also exploits an additional criterion similar to length-based sentence alignment in order to determine whether the segments in corresponding HTML structures have similar lengths. In the current PTMiner, this criterion is not incorporated. However, we have included the sentence-alignment criterion as a later filtering step in Nie and Cai (2001) : If a pair of texts cannot be aligned reasonably well, then that pair is removed. This technique is shown to bring a large improvement for the English-Chinese corpus. A similar approach could also be envisioned for the corpora of European languages, but in the present study, such an approach is not used.", "cite_spans": [ { "start": 100, "end": 113, "text": "Resnik (1998)", "ref_id": "BIBREF40" }, { "start": 447, "end": 465, "text": "Nie and Cai (2001)", "ref_id": "BIBREF33" } ], "ref_spans": [], "eq_spans": [], "section": "Language and Character Set.", "sec_num": "2.6.3" }, { "text": "PTMiner uses heuristics that are mostly language-independent. This allows us to adapt it easily for different language pairs by changing a few parameters (e.g., prefix and suffix lists of file name). It is surprising that so simple an approach is nevertheless very effective. We have been able, using PTMiner, to construct large parallel corpora from the Web for the following language pairs: English-French, English-Italian, English-German, English-Dutch, and English-Chinese. The sizes of these corpora are shown in Table 1 .", "cite_spans": [], "ref_spans": [ { "start": 518, "end": 525, "text": "Table 1", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Mining Results", "sec_num": "2.7" }, { "text": "One question that may be raised is how accurate the mining results are, or how parallel the pages identified are. Actually, it is very difficult to answer this question. We have not undertaken an extensive evaluation but have only performed a simple evaluation with a set of samples. For English-French, from 60 randomly selected candidate sites, AltaVista indexed about 8,000 pages in French. From these, the pair-scanning step identified 4,000 pages with equivalents in English. This showed that the lower bound of recall of pairscanning is 50%. The equivalence of the pair pages identified was judged by an undergraduate student who participated in developing the preliminary version of PTMiner. The criterion used to judge the equivalence of two pages was subjective, with the general guideline being whether two pages describe the same contents and whether they have similar structures. To evaluate precision, 164 pairs of pages from the 4,000 identified were randomly selected and manually checked. It turned out that 162 of them were truly parallel. This shows that the precision is close to 99%. For an English-Chinese corpus, a similar evaluation has been reported in Chen and Nie (2000) . This evaluation was done by a graduate student working on PTMiner. Among 383 pairs randomly selected at the pair-scanning step, 302 pairs were found to be really parallel. The precision ratio is 79%, which is not as good as that of the English-French case. There are several reasons for this:", "cite_spans": [ { "start": 1177, "end": 1196, "text": "Chen and Nie (2000)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Mining Results", "sec_num": "2.7" }, { "text": "\u2022 Incorrect links. It may be that a page is outdated but still indexed by the search engines. A pair including that page will be eliminated in the content-filtering step.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Mining Results", "sec_num": "2.7" }, { "text": "\u2022 Pages that are designed to be parallel, although the contents are not all translated yet. One version of a page may be a simplified version of the other. Some cases of this type can also be filtered out in the content-filtering step, but some will still remain.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Mining Results", "sec_num": "2.7" }, { "text": "\u2022 Pages that are valid parallel pairs yet consist mostly of graphics rather than text. These pages cannot be used for the training of translation models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Mining Results", "sec_num": "2.7" }, { "text": "\u2022 Pairs that are not parallel at all. Filenames of some nonparallel pages may accidentally match the naming rules. For example, . . . /et.html versus . . . /etc.html .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Mining Results", "sec_num": "2.7" }, { "text": "Related to the last reason, we also observed that the names of parallel Chinese and English pages may be very different from one another. For example, it is frequent practice to use the Pinyin translation as the name of a Chinese page of the corresponding English file name (e.g., fangwen.html vs. visit.html ). Another convention is to use numbers as the filenames. For example 1.html would correspond to 2.html . In either of these cases, our pair-scanning approach based on name similarity will fail to recognize the pair. Overall, the naming of Chinese files is much more variable and flexible than the naming of files for European languages. Hence, there exist fewer evident heuristics for Chinese than for the European languages that would allow us to enlarge the coverage and improve the precision of pair scanning.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Mining Results", "sec_num": "2.7" }, { "text": "Given the potentially large number of erroneously identified parallel pairs, a question naturally arises: Can such a noisy corpus actually help CLIR? We will examine this question in Section 4. In the next section we will briefly describe how statistical translation models are trained on parallel corpora. We will focus in our discussion on the following languages: English, French, and Italian. The resulting translation models will be evaluated in a CLIR task.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Mining Results", "sec_num": "2.7" }, { "text": "Bilingual pairs of documents collected from the Web are used as training material for the statistical translation models that we exploit for CLIR. In practice, this material must be organized into a set of small pairs of corresponding segments (typically, sentences), each consisting of a sequence of word tokens. We start by presenting the details of this preparatory step and then discuss the actual construction of the translation models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Building the Translation Models", "sec_num": "3." }, { "text": "3.1 Preparing the Corpus 3.1.1 Format Conversion, Text Segmentation, and Sentence Alignment. The collection process described in the previous section provides us with a set of pairs of HTML files. The first step in preparing this material is to extract the textual data from the files and organize them into small, manageable chunks (sentences).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Building the Translation Models", "sec_num": "3." }, { "text": "In doing so, we try to take advantage of the HTML markup. For instance, we know that

tags normally identify paragraphs,

  • tags mark list items that can also often be interpreted as paragraphs, tags are normally used to mark section headers and may therefore be taken as sentences, and so on.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Building the Translation Models", "sec_num": "3." }, { "text": "Unfortunately, a surprisingly large number of HTML files on the Web are badly formatted, which calls for much flexibility on the part of Web browsers. To help cope with this situation, we employ a freely distributed tool called tidy (Ragget 1998) , which attempts to clean up HTML files, so as to make them XML-compliant. This cleanup process mostly consists in normalizing tag names to the standard XHTML lowercase convention, wrapping tag attributes within double quotes and, most importantly, adding missing tags so as to end up with documents with balancing opening and closing tags.", "cite_spans": [ { "start": 233, "end": 246, "text": "(Ragget 1998)", "ref_id": "BIBREF39" } ], "ref_spans": [], "eq_spans": [], "section": "Building the Translation Models", "sec_num": "3." }, { "text": "Once this cleanup is done, we can parse the files with a standard SGML parser (we use nsgmls [Clark 2001] ) and use the output to produce documents in the standard cesAna format. This SGML format, proposed as part of the Corpus Encoding Standard (CES) (Ide, Priest-Dorman, and V\u00e9ronis 1995) has provisions for annotating simple textual structures such as sections, paragraphs, and sentences. In addition to the cues provided by the HTML tags, we employ a number of heuristics, as well as languagespecific lists of common abbreviations and acronyms, to locate sentence boundaries within paragraphs. When, as sometimes happens, the tidy program fails to make sense of its input on a particular file, we simply remove all SGML markup from the file and treat the document as plain text, which means that we must rely solely on our heuristics to locate paragraph and sentence boundaries.", "cite_spans": [ { "start": 93, "end": 105, "text": "[Clark 2001]", "ref_id": "BIBREF6" }, { "start": 252, "end": 290, "text": "(Ide, Priest-Dorman, and V\u00e9ronis 1995)", "ref_id": "BIBREF12" } ], "ref_spans": [], "eq_spans": [], "section": "Building the Translation Models", "sec_num": "3." }, { "text": "Once the textual data have been extracted from pairs of documents and are neatly segmented into paragraphs and sentences, we can proceed with sentence alignment. This operation produces what we call couples, that is, minimal-size pairs of corresponding segments between two documents. In the vast majority of cases, couples consist of a single pair of sentences that are translations of one another (what we call 1-to-1 couples). However, there are sometimes \"larger\" couples, as when a single sentence in one language translates into two or more sentences in the other language (1-to-N or N-to-1), or when sentences translate many to many (N-to-M). Conversely, there are also \"smaller\" couples, such as when a sentence from either one of the two texts does not appear in the other (0-to-1 or 1-to-0).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Building the Translation Models", "sec_num": "3." }, { "text": "Our sentence alignments are carried out by a program called sfial, an improved implementation of the method described in Simard, Foster, and Isabelle (1992) . For a given pair of documents, this program uses dynamic programming to compute the alignment that globally maximizes a statistical-based scoring function. This function takes into account the statistical distribution of translation patterns (1-to-1, 1-to-N, etc.) and the relative sizes of the aligned text segments, as well as the number of \"cognate\" words within couples, that is, pairs of words with similar orthographies in the two languages (e.g. statistical in English vs. statistique in French).", "cite_spans": [ { "start": 121, "end": 156, "text": "Simard, Foster, and Isabelle (1992)", "ref_id": "BIBREF45" } ], "ref_spans": [], "eq_spans": [], "section": "Building the Translation Models", "sec_num": "3." }, { "text": "The data produced up to this point in the preparation process constitutes what we call a Web-aligned corpus (WAC).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Building the Translation Models", "sec_num": "3." }, { "text": "Since our goal is to use translation models in an IR context, it seems natural to have both the translation models and the IR system operate on the same type of data. The basic indexing units of our IR systems are word stems. Stemming is an IR technique whereby morphologically related word forms are reduced to a common form: a stem. Such a stem does not necessarily have to be a linguistic root form. The principal function of the stem is to serve as an index term in the vocabulary of index terms. Stemming is a form of conflation: Equivalence classes of tokens help to reduce the variance in index terms. Most stemming algorithms fall into two categories: (1) suffix strippers, and (2) full morphological normalization (sometimes referred to as \"linguistic stemming\" in the IR literature). Suffix strippers remove suffixes in an iterative fashion using rudimental morphological knowledge encoded in context-sensitive patterns. The advantage of algorithms of this type (e.g., Porter 1980) is their simplicity and efficiency, although this advantage applies principally to languages with a relatively simple morphology, like English. A different way of generating conflation classes is to employ full morphological analysis. This process usually consists of two steps: First the texts are POS-tagged in order to eliminate each token's part-of-speech ambiguity, and then word forms are reduced to their root form, a process that we refer to as lemmatization. More information about the relative utility of morphological normalization techniques in IR systems can be found in, for example, Hull (1996) , Kraaij and Pohlmann (1996), and Braschler and Ripplinger (2003) .", "cite_spans": [ { "start": 979, "end": 991, "text": "Porter 1980)", "ref_id": "BIBREF38" }, { "start": 1590, "end": 1601, "text": "Hull (1996)", "ref_id": "BIBREF10" }, { "start": 1604, "end": 1614, "text": "Kraaij and", "ref_id": "BIBREF15" }, { "start": 1615, "end": 1635, "text": "Pohlmann (1996), and", "ref_id": "BIBREF16" }, { "start": 1636, "end": 1649, "text": "Braschler and", "ref_id": null }, { "start": 1650, "end": 1667, "text": "Ripplinger (2003)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Tokenization, Lemmatization, and Stopwords.", "sec_num": "3.1.2" }, { "text": "Lemmatizing and removing stopwords from the training material is also beneficial for statistical translation modeling, helping to reduce the problem of data sparseness in the training set. Furthermore, function words and morpho-syntactic features typically arise from grammatical constraints intrinsic to a language, rather than as direct realizations of translated concepts. Therefore, we expect that removing them helps the translation model focus on meaning rather than form. In fact, it has been shown in Chen and Nie (2000) that the removal of stopwords from English-Chinese training material improves both the translation accuracy of the translation models and the effectiveness of CLIR. We expect a similar effect for European languages.", "cite_spans": [ { "start": 509, "end": 528, "text": "Chen and Nie (2000)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Tokenization, Lemmatization, and Stopwords.", "sec_num": "3.1.2" }, { "text": "We also have to tokenize the texts, that is, to identify individual word forms. Because we are dealing with Romance languages, this step is fairly straightforward: 6 We essentially segment the text using blank spaces and punctuation. In addition, we rely on a small number of language-specific rules to deal, for example, with elisions in French (l'amour \u2192 l' + amour) and Italian (dell'arte \u2192 dell' + arte), contractions in French (au \u2192\u00e0 + le), possessives in English (Bob's \u2192 Bob + 's), etc.", "cite_spans": [ { "start": 164, "end": 165, "text": "6", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Tokenization, Lemmatization, and Stopwords.", "sec_num": "3.1.2" }, { "text": "Once we have identified word tokens, we can lemmatize or stem them. For Italian, we relied on a simple, freely distributed stemmer from the Open Muscat project. 7 For French and English, we have access to more sophisticated tools that compute each token's lemma based on its part of speech (we use the HMM-based POS tagger proposed in Foster (1991) and extensive dictionaries with morphological information. As a final step, we remove stopwords.", "cite_spans": [ { "start": 161, "end": 162, "text": "7", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Tokenization, Lemmatization, and Stopwords.", "sec_num": "3.1.2" }, { "text": "Usually, 1-1 alignments are more reliable than other types of alignment. It is a common practice to use only these alignments for model training, and this is what we do. Table 2 provides some statistics on the processed corpora.", "cite_spans": [], "ref_spans": [ { "start": 170, "end": 177, "text": "Table 2", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Tokenization, Lemmatization, and Stopwords.", "sec_num": "3.1.2" }, { "text": "In statistical translation modeling, we take the view that each possible target language text is a potential translation for any given source language text, but that some translations are more likely than others. In the terms of Brown et al. (1990) , a noisy-channel translation model is one that captures this state of affairs in a statistical distribution ", "cite_spans": [ { "start": 229, "end": 248, "text": "Brown et al. (1990)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "P(T | S),", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "where S is a source language text and T is a target language text. 8 With such a model, translating S amounts to finding the target language textT that maximizes P(T | S).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "Modeling P(T | S) is, of course, complicated by the fact that there is an infinite number of possible source and target language texts, and so much of the work of the last 15 years or so in statistical machine translation has been aimed at finding ways to overcome this complexity by making various simplifying assumptions. Typically, P(T | S) is rewritten as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "P(T | S) = P(T)P(S | T) P(S)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "following Bayes' law. This decomposition of P(T | S) is useful in two ways: first, it makes it possible to ignore P(S) when searching forT; second, it allows us to concentrate our efforts on the lexical aspects of P(S | T), leaving it to P(T) (the \"target language model\") to take care of syntactic and other language-specific aspects.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "In one of the simplest and earliest statistical translation models, IBM's Model 1, it is assumed that P(S | T) can be approximated by a computation that uses only \"lexical\" probabilities P(s | t) over source and target language words s and t. In other words, this model completely disregards the order in which the individual words of S and T appear. Although this model is known to be too weak for general translation, it appears that it can be quite useful for an application such as CLIR, because many IR systems also disregard word order, viewing documents and queries as unordered bags of words.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "The P(s | t) distribution is estimated from a corpus of aligned sentences like the one we have produced from our Web-mined collection of bilingual documents, using the expectation maximization (EM) algorithm (Baum 1972) to find the parameters that maximize the likelihood of the training set. As in all machine-learning problems, especially those related to natural language, data sparseness is a critical issue in this process. Even with a large training corpus, many pairs of words (s, t) occur at very low frequencies, and most never occur at all, making it impossible to obtain reliable estimates for the corresponding P(s | t). Without adequate smoothing techniques, lowfrequency events can have disastrous effects on the global behavior of the model, and unfortunately, in natural languages, low-frequency events are the norm rather than the exception.", "cite_spans": [ { "start": 208, "end": 219, "text": "(Baum 1972)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "The goal of translation in CLIR is different from that in general language processing. In the latter case it is important to enable a model to handle low-frequency words and unknown words. For CLIR the coverage of low-frequency words or unknown words by the model is less problematic. Even if a low-frequency word is translated incorrectly, the global IR effectiveness will often not be significantly affected, because low-frequency words likely do not appear often in the document collection to be searched or other terms in the query could compensate for this gap. Most IR algorithms are based on a term-weighting function that favors terms that occur frequently in a document but occur infrequently in the document collection. This means that the best index terms have a medium frequency (Salton and McGill 1983) . Stopwords and (near) hapaxes are less important for IR; limited coverage of very infrequent words in a translation model is therefore not critical for the performance of a CLIR system.", "cite_spans": [ { "start": 791, "end": 815, "text": "(Salton and McGill 1983)", "ref_id": "BIBREF42" } ], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "Proper nouns are special cases of unknown words. When they appear in a query, they usually denote an important part of the user's intention. However, we can adopt a special approach to cope with these unknown words in CLIR without integrating them as the generalized case in the model. For example, one can simply retain all the unknown words in the query translation. This approach works well for most cases in European languages. We have previously shown that a fuzzy-matching approach based on n-grams offers an effective means of overcoming small spelling variations in proper noun spelling (Kraaij, Pohlmann, and Hiemstra 2000) .", "cite_spans": [ { "start": 595, "end": 632, "text": "(Kraaij, Pohlmann, and Hiemstra 2000)", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "The model pruning techniques developed in computational linguistics are also useful for the models used in CLIR. The beneficial effect is that unreliable (or lowprobability) translations can be removed. In Section 4, model smoothing will be motivated from a more theoretical point of view. Here, let us first outline the two variations we used to prune the models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "The first one is simple, yet effective in our application: We consider unreliable all parameters (translation probabilities) whose value falls below some preset threshold (in practice, 0.1 works well). These parameters are simply discarded from the model. The remaining parameters are then renormalized so that all marginal distributions sum to one.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "Another pruning technique is based on the relative contribution to the entropy of the model. We retain the N most reliable parameters (in practice, N = 100K works well). The reliability of a parameter is measured with regard to its contribution to the model's entropy (Foster 2000) . In other words, we discard the parameters that least affect the overall probability of the training set. The remaining parameters are then renormalized so that all marginal distributions sum to one.", "cite_spans": [ { "start": 268, "end": 281, "text": "(Foster 2000)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "Of course, as a result of this, most pairs of words (s, t) are unknown to the translation model (translation probability equals zero). As previously discussed, however, this will not have a disastrous effect on CLIR; on the contrary, some positive effect can be expected as long as there is at least one translation for each source term.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "One important characteristic of these noisy-channel models is that they are \"directional.\" Depending on the intended use, it must be determined beforehand which language is the source and which the target for each pair of languages. Although \"reverse\" parameters can theoretically be obtained from the model through Bayes' rule, it is often more practical to train two separate models if both directions are needed. This topic is also discussed in the next section.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Translation Models", "sec_num": "3.2" }, { "text": "When CLIR is considered simply as a combination of separate MT and IR components, the embedding of the two functions is not a problem. However, as we explained in Section 1, there are theoretical motivations for embedding translation into the retrieval model. Since translation models provide more than one translation, we will try to exploit this extra information, in order to enhance retrieval effectiveness. In Section 4.1 we will first introduce a monolingual probabilistic IR model based on cross entropy between a unigram language model for the query and one for the document. We discuss the relationship of this model to IR models based on generative language models. Subsequently, we show several ways to add translation to the model: One can either translate the query language model from the source language into the target language (i.e., the document language) before measuring the cross entropy, or translate the document model from the target language into the source language and then measure the cross entropy.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Embedding Translation into the IR Model", "sec_num": "4." }, { "text": "Recently, a new approach to IR based on statistical language models has gained wide acceptance. The approach was developed independently by several groups (Ponte and Croft 1998; Miller, Leek, and Schwartz 1999; Hiemstra 1998) and has yielded results on several IR standardized evaluation tasks that are comparable to or better than those obtained using the existing OKAPI probabilistic model. In comparison with the OKAPI model, the IR model based on generative language models has some important advantages: It contains fewer collection-dependent tuning parameters and is easy to extend. For a more detailed discussion of the relationships between the classical (discriminative) probabilistic IR models and recent generative probabilistic IR models, we refer the reader to Kraaij and Spitters (2003) . Probably the most important idea in the language-modeling approach to IR is that documents are scored on the probability that they generate the query; that is, the problem is reversed, an idea that has successfully been applied in speech recognition. There are various reasons that this approach has proven fruitful, probably the most important being that documents contain much more data for estimating the parameters of a probabilistic model than do ad hoc queries (Lafferty and Zhai 2001b) . For ad hoc retrieval, one could describe the query formulation process as follows: A user has an ideal relevant document in mind and tries to describe it by mentioning some of the salient terms that he thinks occur in the document, interspersed with some query stop phrasing like \"Relevant documents mention. . . .\" For each document in the collection, we can compute the probability that the query is generated from a model representing that document. This generation process can serve as a coarse way of modeling the user's query formulation process. The query likelihood given each document can directly be used as a document-ranking function. Formula (1) shows the basic language model, in which a query Q consists of a sequence of terms T 1 , T 2 , . . . , T m that are sampled independently from a document unigram model for document d k (Table 3 presents an explanation of the most important symbols used in equations (1)-(12)):", "cite_spans": [ { "start": 155, "end": 177, "text": "(Ponte and Croft 1998;", "ref_id": "BIBREF37" }, { "start": 178, "end": 210, "text": "Miller, Leek, and Schwartz 1999;", "ref_id": "BIBREF30" }, { "start": 211, "end": 225, "text": "Hiemstra 1998)", "ref_id": null }, { "start": 774, "end": 800, "text": "Kraaij and Spitters (2003)", "ref_id": "BIBREF18" }, { "start": 1270, "end": 1295, "text": "(Lafferty and Zhai 2001b)", "ref_id": "BIBREF22" } ], "ref_spans": [ { "start": 2142, "end": 2150, "text": "(Table 3", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "P(Q | D k ) = P(T 1 , T 2 , . . . , T m | D k ) \u2248 m j=1 P(T m | M D k )", "eq_num": "( 1)" } ], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "In this formula M D k denotes a language model of D k . It is indeed an approximation of D k . Now, if a query is more probable given a language model based on document D 1 than given a language model based on document D 2 , we can then hypothesize that document D 1 is more likely to be relevant to the query than document D 2 . Thus the probability of generating a certain query given a document-based language model can serve as a score for ranking documents with respect to topical relevance. It is common practice to work with log probabilities, which has the advantage of reducing products to summations. We will therefore rewrite (1) in logarithmic form. Since terms might occur more than once in a query, we prefer to work with types \u03c4 i instead of tokens", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "T i . So c(Q, \u03c4 i )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "is the number of occurrences of \u03c4 i in Q (query term frequency); we will also omit the document subscript k in the following presentation:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "log P(Q | D) = n i=1 c(Q, \u03c4 i ) log P(\u03c4 i | M D )", "eq_num": "( 2)" } ], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "A second core technique from speech recognition that plays a vital role in language models for IR is smoothing. One obvious reason for smoothing is to avoid assigning zero probabilities for terms that do not occur in a document because the term probabilities are estimated using maximum-likelihood estimation. 9 If a single query term does not occur in a document, this would result in a zero probability of generating that query, which might not be desirable in many cases, since documents discuss a certain topic using only a finite set of words. It is very well possible that a term that is highly relevant for a particular topic may not appear in a given document, since it is a synonym for other terms that are also highly relevant. Longer documents will in most cases have a better coverage of relevant index terms (and consequently better probability estimates) than short documents, so one could let the level of smoothing depend on the length of the document (e.g., Dirichlet priors). A second reason for smoothing probability estimates of a generative model for queries is that queries consist of (1) terms that have a high probability of occurrence in relevant documents and (2) terms that are merely used to formulate a proper query statement (e.g., \"Documents discussing only X are not relevant\"). A mixture of a document language model and a language model of typical query terminology (estimated on millions of queries) would probably give good results (in terms of a low perplexity).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "We have opted for a simple approach that addresses both issues, namely, applying a smoothing step based on linear interpolation with a background model estimated on a large document collection, since we do not have a collection of millions of queries:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "log P(Q | D) = n i=1 c(Q, \u03c4 i ) log((1 \u2212 \u03bb)P(\u03c4 i | M D ) + \u03bbP(\u03c4 i | M C ))", "eq_num": "(3)" } ], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "Here, P(\u03c4 i | M C ) denotes the marginal probability of observing the term \u03c4 i , which can be estimated on a large background corpus, and \u03bb is the smoothing parameter. A common range for \u03bb is 0.5-0.7, which means that document models have to be smoothed quite heavily for optimal performance. We hypothesize that this is mainly due to the querymodeling role of smoothing. Linear interpolation with a background model has been frequently used to smooth document models (e.g., Miller, Leek, and Schwartz 1999; Hiemstra 1998) . Recently other smoothing techniques (Dirichlet, absolute discounting) have also been evaluated. An initial attempt to account for the two needs for smoothing (sparse data problem, query modeling) with separate specialized smoothing functions yielded positive results (Zhai and Lafferty 2002) .", "cite_spans": [ { "start": 475, "end": 507, "text": "Miller, Leek, and Schwartz 1999;", "ref_id": "BIBREF30" }, { "start": 508, "end": 522, "text": "Hiemstra 1998)", "ref_id": null }, { "start": 792, "end": 816, "text": "(Zhai and Lafferty 2002)", "ref_id": "BIBREF52" } ], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "We have tested the model corresponding to formula (3) in several different IR applications: monolingual information retrieval, filtering, topic detection, and topic tracking (cf. Allen [2002] for a task description of the latter two tasks). For several of these applications (topic tracking, topic detection, collection fusion), it is important counts of x that scores be comparable across different queries (Spitters and Kraaij 2001 ). The basic model does not provide such comparability of scores, so it has to be extended with score normalization. There are two important steps in doing this. First of all, we would like to normalize across query specificity. The generative model will produce low scores for specific queries (since the average probability of occurrence is low) and higher scores for more general queries. Normalization can be accomplished by modeling the IR task as a likelihood ratio (Ng 2000) . For each term in the query, the log-likelihood ratio (LLR) model judges how surprising it is to see the term, given the document model in comparison with the background model:", "cite_spans": [ { "start": 179, "end": 191, "text": "Allen [2002]", "ref_id": "BIBREF0" }, { "start": 408, "end": 433, "text": "(Spitters and Kraaij 2001", "ref_id": "BIBREF46" }, { "start": 906, "end": 915, "text": "(Ng 2000)", "ref_id": "BIBREF31" } ], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "LLR(Q | D) = log P(Q | M D ) P(Q | M C ) = n i=1 c(Q, \u03c4 i ) log ((1 \u2212 \u03bb)P(\u03c4 i | M D ) + \u03bbP(\u03c4 i | M C )) P(\u03c4 i | M C )", "eq_num": "(4)" } ], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "In (4), P(Q | M C ) denotes the generative probability of the query given a language model estimated on a large background corpus C. Note that P(Q | M C ) is a querydependent constant and does not affect document ranking. Actually, model (4) has a better justification than model (3), since it can be seen as a direct derivative of the log-odds of relevance if we assume uniform priors for document relevance:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "log P(R | D, Q) P(R | D, Q) = log P(Q | R, D) P(Q |R, D) + log P(R | D) P(R | D) \u2248 log P(Q | M D ) P(Q | M C ) + K", "eq_num": "(5)" } ], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "In (5), R refers to the event that a user likes a particular document (i.e., the document is relevant). The scores of model (4) still depend on the query length, which can be easily normalized by dividing the scores by the query length ( i c(Q, \u03c4 i )). This results in formula (6) for the normalized log-likelihood ratio (NLLR) of the query:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "NLLR(Q | D) = n i=1 c(Q, \u03c4 i ) i c(Q, \u03c4 i ) log ((1 \u2212 \u03bb)P(\u03c4 i | M D ) + \u03bbP(\u03c4 i | M C )) P(\u03c4 i | M C )", "eq_num": "(6)" } ], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "A next step is to view the normalized query term counts c(Q, \u03c4 i )/ i c(Q, \u03c4 i ) as maximum-likelihood estimates of a probability distribution representing the query P(\u03c4 i | M Q ). The NLLR formula can now be reinterpreted as a relationship between the two probability distributions P(\u03c4 | M Q ), P(\u03c4 | M D ) normalized by the the third distribution P(\u03c4 | M C ). The model measures how much better than the background model the document model can encode events from the query model; or in information-theoretic terms, it can be interpreted as the difference between two cross entropies:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "NLLR(Q | D) = n i=1 P(\u03c4 i | Q) log P(\u03c4 i | D k ) P(\u03c4 i | C) = H(X | c) \u2212 H(X | d)", "eq_num": "( 7)" } ], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "In 7, X is a random variable with the probability distribution p(\u03c4 i ) = p(\u03c4 i | M Q ), and c and d are probability mass functions representing the marginal distribution and the document model. Cross entropy is a measure of our average surprise, so the better a document model \"fits\" a particular query distribution, the higher its score will be. 10 The representation of both the query and a document as samples from a distribution representing, respectively, the user's request and the document author's \"mindset\" has several advantages. Traditional IR techniques like query expansion and relevance feedback can be reinterpreted in an intuitive framework of probability distributions (Lafferty and Zhai 2001a; Lavrenko and Croft 2001) . The framework also seems suitable for cross-language retrieval. We need only to extend the model with a translation function, which relates the probability distribution in one language to the probability distribution function in another language. We will present several solutions for this extension in the next section.", "cite_spans": [ { "start": 347, "end": 349, "text": "10", "ref_id": null }, { "start": 686, "end": 711, "text": "(Lafferty and Zhai 2001a;", "ref_id": "BIBREF21" }, { "start": 712, "end": 736, "text": "Lavrenko and Croft 2001)", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "The NLLR also has a disadvantage: It is less easy in the NLLR to integrate prior information about relevance into the model (Kraaij, Westerveld, and Hiemstra 2002) , which can be done in a straightforward way in formula (1), by simple multiplication. CLIR is a special case of ad hoc retrieval, and usually a document length-based prior can enhance results significantly. A remedy that has proven to be effective is linear interpolation of the NLLR score with a prior log-odds ratio log (P(R | D)/P(\u00acR | D) (Kraaij 2002) . For reasons of clarity, we have chosen not to include this technique in the experiments presented here.", "cite_spans": [ { "start": 124, "end": 163, "text": "(Kraaij, Westerveld, and Hiemstra 2002)", "ref_id": "BIBREF19" }, { "start": 507, "end": 520, "text": "(Kraaij 2002)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "In the following sections, we will describe several ways to extend the monolingual IR model with translation. The section headings include the run tags that will be used in Section 5 to describe the experimental results.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual IR Based on Unigram Language Models", "sec_num": "4.1" }, { "text": "In Section 4.1, we have seen that the basic retrieval model measures the cross entropy between two language models: a language model of the query and a language model of the document. 11 Instead of translating a query before estimating a query model (the external approach), we propose to estimate the query model directly in the target language. We will do this by decomposing the problem into two components that are easier to estimate:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Estimating the Query Model in the Target Language (QT)", "sec_num": "4.2" }, { "text": "P(t i | M Qs ) = L j P(s j , t i | M Qs ) = L j P(t i | s j , M Qs )P(s j | M Qs ) \u2248 L j P(t i | s j )P(s j | M Qs ) (8)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Estimating the Query Model in the Target Language (QT)", "sec_num": "4.2" }, { "text": "where L is the size of the source vocabulary. Thus, P(t i | M Qs ) can be approximated by combining the translation model P(t i | s j ), which we can estimate on the parallel Web corpus, and the familiar P(s j | M Qs ), which can be estimated using relative frequencies. This simplified model, from which we have dropped the dependency of P(t i | s j ) on Q, can be interpreted as a way of mapping the probability distribution function in the source language event space P(s j | M Qs ) onto the event space of the target language vocabulary. Since this probabilistic mapping function involves a summation over all possible translations, mapping the query model from the source language can be implemented as the matrix product of a vector representing the query probability distribution over source language terms with the translation matrix P(t i | s j ). 12 The result is a probability distribution function over the target language vocabulary.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Estimating the Query Model in the Target Language (QT)", "sec_num": "4.2" }, { "text": "Now we can substitute the query model P(\u03c4 i | M Q ) in formula (7) with the target language query model in (8) and, after a similar substitution operation for P(\u03c4 i | M C ), we arrive at CLIR model QT:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Estimating the Query Model in the Target Language (QT)", "sec_num": "4.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "NLLR-QT(Q s | D t ) = n i=1 L j=1 P(t i | s j )P(s j | M Qs ) log (1 \u2212 \u03bb)P(t i | M Dt ) + \u03bbP(t i | M Ct ) P(t i | M Ct )", "eq_num": "(9)" } ], "section": "Estimating the Query Model in the Target Language (QT)", "sec_num": "4.2" }, { "text": "Another way to embed translation into the IR model is to estimate the document model in the source language:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Estimating the Document Model in the Source Language (DT)", "sec_num": "4.3" }, { "text": "P(s i | M Dt ) = N j P(s i , t j | M Dt ) = N j P(s i | t j , M Dt )P(t j | M Dt ) \u2248 N j P(s i | t j )P(t j | M Dt ) (10)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Estimating the Document Model in the Source Language (DT)", "sec_num": "4.3" }, { "text": "where N is the size of the target vocabulary. Obviously, we need a translation model in the reverse direction for this approach. Now we can substitute (10) for P(\u03c4 i | M D ) in formula (6), yielding CLIR model DT:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Estimating the Document Model in the Source Language (DT)", "sec_num": "4.3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "NLLR-DT(Q s | D t ) = n i=1 P(s i | M Qs ) log N j=1 P(s i | t j )((1 \u2212 \u03bb)P(t j | M Dt ) + \u03bbP(t j | M Ct )) N j=1 P(s i | t j )P(t j | M Ct )", "eq_num": "(11" } ], "section": "Estimating the Document Model in the Source Language (DT)", "sec_num": "4.3" }, { "text": ") It is important to realize that both the QT and DT models are based on contextinsensitive translation, since translation is added to the IR model after the independence assumption (1) has been made. Recently, a more complex CLIR model based on relaxed assumptions-context-sensitive translation but term independence-based IRhas been proposed in Federico and Bertoldi (2002) . In experiments on the CLEF test collections, the aforementioned model also proved to be more effective; however, it has the disadvantage of reducing efficiency through its use of a Viterbi search procedure.", "cite_spans": [ { "start": 347, "end": 375, "text": "Federico and Bertoldi (2002)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Estimating the Document Model in the Source Language (DT)", "sec_num": "4.3" }, { "text": "In this section we will discuss several variant instantiations of the QT and DT models that help us measure the importance of the number of translations (pruning) and the weighting of translation alternatives. We also present several baseline CLIR algorithms taken from the literature and discuss their relationship to the QT and DT models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Variant Models and Baselines", "sec_num": "4.4" }, { "text": "As we argued in Section 1, the most simple solution to CLIR is to use an MT system to translate the query and use the translation as the basis for a monolingual search operation in the target language. This solution does not require any modification to the standard IR model as presented in Section 4.1. We will refer to this model as the external-translation approach. The translated query is used to estimate a probability distribution for the query in the target language. Thus, the order of operations is: (1) translate the query using an external tool; (2) estimate the parameters P(t i | M Qt ) of a language model based on this translated query.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "External Translation (MT, NAIVE).", "sec_num": "4.4.1" }, { "text": "In our experimental section below, we will list results with two different instantiations of the external-translation approach: (1) MT: query translation by Systran, which attempts to use high-level linguistic analysis, context-sensitive translation, extensive dictionaries, etc., and (2) NAIVE: naive replacement of each query term with its translations (not weighted). The latter approach is often implemented using bilingual word lists for CLIR. It is clear that this approach can be problematic for terms with many translations, since they would then be assigned a higher relative importance. The NAIVE method is included here only to study the effect of the number of translations on the effectiveness of various models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "External Translation (MT, NAIVE).", "sec_num": "4.4.1" }, { "text": "In Section 3.2 we explained that there are different possible strategies for pruning the translation model. An extreme pruning method is best match, in which only the best translation is kept. A best-match translation model for query model translation (QT-BM) could also be viewed as an instance of the external translation model, but one that uses a corpus-based disambiguation method. Each query term is translated by the most frequent translation in the Web corpus, disregarding the query context.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Best-Match Translation (QT-BM).", "sec_num": "4.4.2" }, { "text": "If we don't know the precise probability of each translation alternative for a given term, the best thing to do is to fall back on uniform translation probabilities. This situation arises, for example, if we have only standard bilingual dictionaries. We hypothesize that this approach will be more effective than NAIVE but less effective than QT.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Equal Probabilities (QT-EQ).", "sec_num": "4.4.3" }, { "text": "An alternative way to embed translation into the retrieval model is to view translation alternatives as synonyms. This is, of course, something of an idealization, yet there is certainly some truth to the approach when translations are looked up in a standard bilingual dictionary. Strictly speaking, when terms are pure synonyms, they can be substituted for one another. Combining translation alternatives with the synonym operator of the INQUERY IR system (Broglio et al. 1995) , which conflates terms on the fly, has been shown to be an effective way of improving the performance of dictionary-based CLIR systems (Pirkola 1998) . In our study of stemming algorithms (Kraaij and Pohlmann 1996) , we independently implemented the synonym operator in our system. This on-line conflation function replaces the members of the equivalence class with a class ID, usually a morphological root form. We have used this function to test the effectiveness of a synonymy-based CLIR model in a language model IR setting.", "cite_spans": [ { "start": 458, "end": 479, "text": "(Broglio et al. 1995)", "ref_id": "BIBREF3" }, { "start": 616, "end": 630, "text": "(Pirkola 1998)", "ref_id": "BIBREF36" }, { "start": 669, "end": 695, "text": "(Kraaij and Pohlmann 1996)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Synonym-Based Translation (SYN).", "sec_num": "4.4.4" }, { "text": "The synonym operator for CLIR can be formalized as the following class equivalence model (assuming n translations t j for term s i and N unique terms in the target language):", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Synonym-Based Translation (SYN).", "sec_num": "4.4.4" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "P(class(s i ) | M Dt ) = n j c(t j , D t ) N j c(t j , D t ) = N j \u03b4(s i , t j )P(t j | M Dt )", "eq_num": "( 12)" } ], "section": "Synonym-Based Translation (SYN).", "sec_num": "4.4.4" }, { "text": "where", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Synonym-Based Translation (SYN).", "sec_num": "4.4.4" }, { "text": "P(class(s i ) | M Dt )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Synonym-Based Translation (SYN).", "sec_num": "4.4.4" }, { "text": "is the probability that a member of the equivalence class of s i is generated by the language model M Dt and", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Synonym-Based Translation (SYN).", "sec_num": "4.4.4" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "\u03b4(s i , t j ) = 1 if t j \u2208 class(s i ) 0 if t j / \u2208 class(s i )", "eq_num": "(13)" } ], "section": "Synonym-Based Translation (SYN).", "sec_num": "4.4.4" }, { "text": "Here c(t j , D t ) is the term frequency (counts) of term t j in document D t . The synonym class function \u03b4(s i , t j ) can be interpreted as a special instantiation of the translation model P(s i | t j ) in (10), namely, P(s i | t j ) = 1 for all translations t j of s i . Of course, this does not yield a valid probability function, since the translation probabilities for all translations s i of a certain t j do not sum to one, because the pseudosynonym classes are not disjunct because of sense ambiguity. But the point is that the structure of a probabilistic version of the SYN model is similar to that of the DT model, namely, one in which all translations have a reverse translation probability P(s i | t j ) equal to one. This is obviously just an approximation of reality. We therefore expect that this model will be less effective than the QT and DT models. In our implementation of the SYN model, we formed equivalence classes by looking up all translations of a source term s i in the translation model P(t j | s i ). The translations receive a weight of one and are used as pseudo translation-probabilities 13 in the model corresponding to formula (11).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Synonym-Based Translation (SYN).", "sec_num": "4.4.4" }, { "text": "In dictionary-based approaches, the number of translation alternatives is usually not as high as in (unpruned) translation models, so these alternatives can be used in some form of query expansion (Hull and Grefenstette 1996; Savoy 2002) . However, it is well known that most IR models break down when the number of translations is high. To remedy this, researchers have tried to impose query structure, for example, by collecting translation alternatives in an equivalence class (Pirkola 1998) , or via a quasi-Boolean structure (Hull 1997) .", "cite_spans": [ { "start": 197, "end": 225, "text": "(Hull and Grefenstette 1996;", "ref_id": "BIBREF10" }, { "start": 226, "end": 237, "text": "Savoy 2002)", "ref_id": "BIBREF43" }, { "start": 480, "end": 494, "text": "(Pirkola 1998)", "ref_id": "BIBREF36" }, { "start": 530, "end": 541, "text": "(Hull 1997)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "4.5" }, { "text": "The idea of embedding a translation step into an IR model based on query likelihood was developed independently by several researchers (Hiemstra and de Jong 1999; Kraaij, Pohlmann, and Hiemstra 2000; Berger and Lafferty 2000) . Initially translation probabilities were estimated from machine-readable dictionaries, using simple heuristics (Hiemstra et al. 2001) . Other researchers have successfully used models similar to DT, in combination with translation models trained on parallel corpora, though not from the Web (McNamee and Mayfield 2001; Xu, Weischedel, and Nguyen 2001) .", "cite_spans": [ { "start": 163, "end": 199, "text": "Kraaij, Pohlmann, and Hiemstra 2000;", "ref_id": "BIBREF17" }, { "start": 200, "end": 225, "text": "Berger and Lafferty 2000)", "ref_id": "BIBREF2" }, { "start": 339, "end": 361, "text": "(Hiemstra et al. 2001)", "ref_id": null }, { "start": 519, "end": 546, "text": "(McNamee and Mayfield 2001;", "ref_id": "BIBREF28" }, { "start": 547, "end": 579, "text": "Xu, Weischedel, and Nguyen 2001)", "ref_id": "BIBREF50" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "4.5" }, { "text": "We carried out several contrastive experiments to gain more insight into the relative effectiveness of the various CLIR models presented in Sections 4.2-4.4. We will first outline our research questions, before describing the experiments in more detail.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiments", "sec_num": "5." }, { "text": "The research questions we are hoping to answer are the following:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Research Questions", "sec_num": "5.1" }, { "text": "How do CLIR systems based on translation models perform with respect to reference systems (e.g., monolingual, MT )?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "1.", "sec_num": null }, { "text": "2. Which manner of embedding a translation model is most effective for CLIR? How does a probabilistically motivated embedding compare with a synonym-based embedding?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "1.", "sec_num": null }, { "text": "3. Is there a query expansion effect, and if so, how can we exploit it?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "1.", "sec_num": null }, { "text": "What is the relative importance of pruning versus weighting?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4.", "sec_num": null }, { "text": "5. Which models are robust against noisy translations?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4.", "sec_num": null }, { "text": "The first two questions concern the main goal of our experiments: What is the effectiveness of a probabilistic CLIR system in which translation models mined from the Web are an integral part of the model, compared to that of CLIR models in which translation is merely an external component? The remaining questions help us to understand the relative importance of various design choices in our approach, such as pruning and translation model orientation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4.", "sec_num": null }, { "text": "We have defined a set of contrastive experiments in order to help us answer the research questions presented above. These experiments seek to compare:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "5.2" }, { "text": "1. The effectiveness of approaches incorporating a translation model produced from the Web to that of a monolingual baseline and an off-the-shelf external query translation approach based on Systran (MT).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "5.2" }, { "text": "The effectiveness of embedding query model translation (QT) and that of document model translation (DT).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "2.", "sec_num": null }, { "text": "The effectiveness of using the entire set of translations, each of which is weighted, (QT) to that of using just the most probable translation (QT-BM).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "3.", "sec_num": null }, { "text": "The effectiveness of weighted query model translation (QT) to that of equally weighted translations (QT-EQ) and nonweighted translations (NAIVE).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "4.", "sec_num": null }, { "text": "The effectiveness of treating translations as synonyms (SYN) with that of weighted translations (QT) and equally weighted translations (QT-EQ).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "5.", "sec_num": null }, { "text": "6. Different translation model pruning strategies: best N parameters or thresholding probabilities.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "5.", "sec_num": null }, { "text": "Each strategy is represented by a run tag, as shown in Table 4 . Table 5 illustrates the differences among the different translation methods. It lists, for several CLIR models, the French translations of the word drug taken from one of the test queries that talks about drug policy.", "cite_spans": [], "ref_spans": [ { "start": 55, "end": 62, "text": "Table 4", "ref_id": "TABREF3" }, { "start": 65, "end": 72, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "5.", "sec_num": null }, { "text": "The translations in Table 5 are provided by the translation models P(e | f ) and P(f | e). The translation models have been pruned by discarding the translations with P < 0.1 and renormalizing the model (except for SYN), or by retaining the 100K best parameters of the translation model. The first pruning method (probability threshold) has a very different effect on the DT method than on the QT method: The number of terms that translate into drug, according to P(e | f ), is much larger than the number of translations of drug found in P(f | e). There are several possible explanations for this: Quite a few French terms, including the verb droguer and the compounds pharmacor\u00e9sistance and pharmacoth\u00e9rapie, all translate into an English expression or compound Table 5 Example translations: Stems and probabilities with different CLIR methods.", "cite_spans": [], "ref_spans": [ { "start": 20, "end": 27, "text": "Table 5", "ref_id": null }, { "start": 764, "end": 771, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "5.", "sec_num": null }, { "text": "MT drogues QT P(f | e) \u2264 0.1 QT-EQ QT-BM SYN NAIVE ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Run ID Translation Translation Model", "sec_num": null }, { "text": "DT P(e | f ) \u2264 0.1 QT ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Run ID Translation Translation Model", "sec_num": null }, { "text": "P(e | f ), 100K", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Run ID Translation Translation Model", "sec_num": null }, { "text": "DT ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Run ID Translation Translation Model", "sec_num": null }, { "text": "P(f | e), 100K", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Run ID Translation Translation Model", "sec_num": null }, { "text": "involving the word drug. Since our translation model is quite simple, these compoundcompound translations are not learned. 14 A second factor that might play a role is the greater verbosity of French texts compared to their English equivalents (cf. Table 2 ). For the models that have been pruned using the 100K-best-parameters criterion, the differences between QT and DT are smaller. Both methods yield multiple translations, most of which seem related to drug, so there is a clear potential for improved recall as a result of the query expansion effect. Notice, however, that the expansion concerns both the medical and the narcotic senses of the word drug. We will see in the following section that the CLIR model is able to take advantage of this query expansion effect, even if the expansion set is noisy and not disambiguated.", "cite_spans": [], "ref_spans": [ { "start": 249, "end": 256, "text": "Table 2", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Run ID Translation Translation Model", "sec_num": null }, { "text": "To answer the research questions stated in section 5.1, we carried out a series of experiments on a combination of the CLEF-2000 CLEF- , -2001 CLEF- and -2002 This joint test collection consists of documents in several languages (articles from major European newspapers from the year 1994), 140 topics describing different information needs (also in several languages) and their corresponding relevance judgments. (Relevance judgments are a human-produced resource that states, for a subset of a document collection, whether a document is relevant for a particular query.) We used only the English, Italian, and French data for the CLIR experiments reported here. The main reason for this limitation was that the IR experiments and translation models were developed at two different sites equipped with different proprietary tools. We chose language pairs for which the lemmatization/stemming step for both the translation model training and indexing system were equivalent. A single test collection was created by merging the three topic sets in order to increase the reliability of our results and sensitivity of significance tests. Each CLEF topic consists of three parts: title, description, and narrative. An example is given below:", "cite_spans": [ { "start": 119, "end": 128, "text": "CLEF-2000", "ref_id": null }, { "start": 129, "end": 142, "text": "CLEF- , -2001", "ref_id": null }, { "start": 143, "end": 158, "text": "CLEF- and -2002", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "The CLEF Test Collection", "sec_num": "5.3" }, { "text": " C001", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The CLEF Test Collection", "sec_num": "5.3" }, { "text": " Architecture in Berlin <description> Find documents on architecture in Berlin.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The CLEF Test Collection", "sec_num": "5.3" }, { "text": "<narrative> Relevant documents report, in general, on the architectural features of Berlin or, in particular, on the reconstruction of some parts of the city after the fall of the Wall.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The CLEF Test Collection", "sec_num": "5.3" }, { "text": "We used only the title and description parts of the topics and concatenated these simply to form the queries. Table 6 lists some statistics on the test collection. 16 The documents were submitted to the preprocessing (stemming/lemmatization) procedure we described in Section 3.1.2. However, for English and French lemmatization, we used the Xelda tools from XRCE, 17 which perform morphological normalization slightly differently from the one described in Section 3.1.2. However, since the two lemmatization strategies are based on the same principle (POS tagging plus inflection removal), the small differences in morphological dictionaries and POS tagging had no significant influence on retrieval effectiveness. 18 All runs use a smoothing parameter \u03bb = 0.3. This value had been shown to work well for CLIR experiments with several other collections.", "cite_spans": [ { "start": 164, "end": 166, "text": "16", "ref_id": null }, { "start": 716, "end": 718, "text": "18", "ref_id": null } ], "ref_spans": [ { "start": 110, "end": 117, "text": "Table 6", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "The CLEF Test Collection", "sec_num": "5.3" }, { "text": "The effectiveness of retrieval systems can be evaluated using several measures. The basic measures are precision and recall, which cannot be applied directly, since they assume clearly separated classes of relevant and nonrelevant documents. The most widely accepted measure for evaluating effectiveness of ranked retrieval systems is the average uninterpolated precision, most often referred to as mean average precision (MAP), since the measure is averaged first over relevant documents and then across topics. Other measures, such as precision at a fixed rank, interpolated precision, or R-precision, are strongly correlated to the mean average precision, so they do not really provide additional information (Tague-Sutcliffe and Blustein 1995; Voorhees 1998).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Measuring Retrieval Effectiveness", "sec_num": "5.4" }, { "text": "The average uninterpolated precision for a given query and a given system version can be computed as follows: First identify the rank number n of each relevant document in a retrieval run. The corresponding precision at this rank number is defined as the number of relevant documents found in the ranks equal to or higher than the respective rank r divided by n. Relevant documents that are not retrieved are assigned a precision of zero. The average precision for a given query is defined as the average value of the precision pr over all known relevant documents d ij for that query. Finally, the mean average precision can be calculated by averaging the average precision over all M queries:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Measuring Retrieval Effectiveness", "sec_num": "5.4" }, { "text": "MAP = 1 M M j=1 1 N j N j i=1 pr(d ij ), where pr(d ij ) = rn i n i , if d ij retrieved and n i \u2264 C 0,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Measuring Retrieval Effectiveness", "sec_num": "5.4" }, { "text": "in other cases (14) Here, n i denotes the rank of the document d ij , which has been retrieved and is relevant for query j, r n i is the number of relevant documents found up to and including rank n i , N j is the total number of relevant documents of query j, M is the total number of queries, and C is the cutoff rank (C is 1,000 for TREC experiments).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Measuring Retrieval Effectiveness", "sec_num": "5.4" }, { "text": "Since we compared many different system versions, which do not always display a large difference in effectiveness, it is desirable to perform significance tests on the results. However, it is well known that parametric tests for data resulting from IR experiments are not very reliable, since the assumptions of these tests (normal or symmetric distribution, homogeneity of variances) are usually not met. We checked the assumptions for an analysis of variance (by fitting a linear model for a withinsubjects design) and found that indeed the distribution of the residual error was quite skewed, even after transformation of the data. Therefore, we resorted to a nonparametric alternative for the analysis of variance, the Friedman test (Conover 1980) . This test is preferable, for the analysis of groups of runs instead, to multiple sign-tests or Wilcoxon signed-rank tests, since it provides overall alpha protection. This means that we first test whether there is any significant difference at all between the runs, before applying multiple-comparison tests. Applying just a large number of paired significance tests at the 0.05 significance level without a global test leads very quickly to a high overall alpha. After applying the Friedman test, we ran Fisher's LSD multiplecomparison tests (recommended by Hull) to identify equivalence classes of runs (Hull 1993; Hull, Kantor, and Ng 1999) . An equivalence class is a group of runs that do not differ significantly (e.g., in terms of mean average precision) from one another in terms of performance.", "cite_spans": [ { "start": 737, "end": 751, "text": "(Conover 1980)", "ref_id": "BIBREF7" }, { "start": 1359, "end": 1370, "text": "(Hull 1993;", "ref_id": null }, { "start": 1371, "end": 1397, "text": "Hull, Kantor, and Ng 1999)", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Measuring Retrieval Effectiveness", "sec_num": "5.4" }, { "text": "We decided to have two types of baseline runs. It is standard practice to take a monolingual run as a baseline. This run is based on an IR system using document ranking formula (6). Contrary to runs described in Kraaij (2002) , we did not use any additional performance-enhancing devices, like document length-based priors or fuzzy matching, in order to focus on the basic retrieval model extensions, avoiding interactions.", "cite_spans": [ { "start": 212, "end": 225, "text": "Kraaij (2002)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Baseline Systems", "sec_num": "5.5" }, { "text": "Systran was used as an additional cross-language baseline, to serve as a reference point for cross-language runs. Notice that the lexical coverage of MT systems varies considerably across language pairs; in particular, the French-English version of Systran is quite good in comparison with those available for other language pairs. We accessed the Web-based version of Systran (December 2002), marketed as Babelfish, using the Perl utility babelfish.pm and converted the Unicode output to the ISO-Latin1 character set to make it compatible with the Xelda-based morphology. Table 7 shows the results for the different experimental conditions in combination with a translation model pruned with the probability threshold criterion P > 0.1 (cf. Section 3.2). For each run, we computed the mean average precision using the standard evaluation tool trec eval. We performed Friedman tests on all the runs based on the Web translation models, because these are the runs in which we are most interested; furthermore, one should avoid adding runs that are quite different to a group that is relatively homogeneous, since this can easily lead to a false global-significance test. The Friedman test (as measured on the F distribution) proved significant at the p < 0.05 level in all cases, so we created equivalence classes using Fisher's LSD method, which are denoted by letters. Letters are assigned to classes in decreasing order of performance; so if a run is a member of equivalence class a, it is one of the best runs for that particular experimental condition.", "cite_spans": [], "ref_spans": [ { "start": 573, "end": 580, "text": "Table 7", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Baseline Systems", "sec_num": "5.5" }, { "text": "The last four rows of the table provide some additional statistics on the query translation process. For both the forward (P(t | s),fw) and the reverse (P(s | t),rev) translation model, we list the percentage of missed translations 19 of unique query terms and the average number of translations per unique query term. Table 8 shows the results for the same experimental conditions, but this time the translation models were pruned by taking the n best translation relations according to an entropy criterion, where n = 100, 000. Several other similar pruning methods have also been tested on the CLEF-2000 subset of the data (e.g. P > 0.01, P > 0.05, 1M parameters, 10K parameters). However, the two cases shown in Tables 7 and 8 represent the best of the two families of pruning techniques. Our goal was not to do extensive parameter tuning in order to find the best-performing combination of models, but rather to detect some broad characteristics of the pruning methods and their interactions with the retrieval model. Since the pruned forward and reverse translation models yield different translation relations (cf. Table 5), we hypothesized that it might be effective to combine them. Instead of combining the translation probabilities directly, we chose to combine the results of the QT and DT models by interpolation of the document scores. Results for combinations based on the 100K models are shown in Table 9 . Indeed, for all the language pairs, the combination run improves upon each of its component runs. This means that each component run can compensate for missing translations in the companion translation model. 5.7 Discussion 5.7.1 Web-Based CLIR versus MT-Based CLIR. Our first observation when examining the data is that the runs based on translation models are comparable to or better than the Systran run. Sign tests showed that there was no significant difference between the MT and QT runs for English-French and French-English language pairs. The QT runs were significantly better at the p = 0.01 level for the Italian-English and English-Italian language pairs. This is a very significant result, particularly since the performance of CLIR with Systran has often been among the best in the previous CLIR experiments in TREC and CLEF. These results show that the Web-based translation models are effective means for CLIR tasks. The better results obtained with the Web-based translation models confirm our intuition, stated in Section 1, that there are better tools for query translation in CLIR than off-the-shelf commercial MT systems.", "cite_spans": [], "ref_spans": [ { "start": 319, "end": 326, "text": "Table 8", "ref_id": null }, { "start": 1413, "end": 1420, "text": "Table 9", "ref_id": "TABREF6" } ], "eq_spans": [], "section": "Results", "sec_num": "5.6" }, { "text": "Compared to the monolingual runs, the best CLIR performance with Web-based translation models varies from 74.1% to 93.7% (80% to 96% for the combined QT+DT models) of the monolingual run. This is within the typical range of CLIR performance. More generally, this research successfully demonstrates the enormous potential of parallel Web pages and Web-based MT.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results", "sec_num": "5.6" }, { "text": "We cannot really compare performance across target languages, since the relevant documents are not distributed in a balanced way: Some queries do not yield any relevant document in some languages. This partly explains why the retrieval effectiveness of the monolingual Italian-Italian run is much higher than the monolingual French and English runs. We can, however, compare methods within a given language pair.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results", "sec_num": "5.6" }, { "text": "Our second question in Section 5.1 concerned the relative effectiveness of the QT and DT models. The experimental results show that there is no clear winner; differences are small and not significant. There seems to be some correlation with translation direction, however: The QT models perform better than DT on the X-English pairs, and the DT models perform better on the English-X pairs. This might indicate that the P(e | f ) and P(e | i) translation models are more reliable than their reverse counterparts. A possible explanation for this could be that the average English sentence is shorter than the corresponding French and Italian sentence. The average number of tokens per sentence is, respectively, 6.6/6.9 and 5.9/6.9 for English/French and English/Italian corpora. This may lead to more reliable estimates for P(e | f ) and P(e | i) than the reverse. However, further investigation is needed to confirm this, since differences in morphology could also contribute to the observed effect. Still, the fact that QT models perform just as well as DT models in combination with translation models is a new result.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Comparison of Query Model Translation (QT), Document Model Translation (DT), and Translations Modeled as Synonyms (SYN).", "sec_num": "5.7.2" }, { "text": "We also compared our QT and DT models to the synonym-based approach (SYN) (Pirkola 1998) . Both the QT and DT models were significantly more effective than the synonym-based model. The latter seems to work well when the number of translations is relatively small but cannot effectively handle the large number of (pseudo)translations produced by our 100K translation models. The synonym-based model usually performs better than the models based on query translation with uniform probabilities, but the differences are not significant in most cases.", "cite_spans": [ { "start": 74, "end": 88, "text": "(Pirkola 1998)", "ref_id": "BIBREF36" } ], "ref_spans": [], "eq_spans": [], "section": "Comparison of Query Model Translation (QT), Document Model Translation (DT), and Translations Modeled as Synonyms (SYN).", "sec_num": "5.7.2" }, { "text": "In Section 1 we argued that using just one translation (as MT does) is probably a suboptimal strategy for CLIR, since there is usually more than one good translation for a particular term. Looking at probabilistic dictionaries, we have also seen that the distinction between a translation and a closely related term cannot really be made on the basis of some thresholding criterion. Since it is well known in IR that adding closely related terms can improve retrieval effectiveness, we hypothesized that adding more than one translation would also help. The experimental results confirm this effect. In all but one case (English-French, P > 0.1), using all translations (QT) yielded significantly better performance than choosing just the most probable translation (QT-BM). For the P > 0.1 models, the average number of translations in the forward direction is only 1.65, so the potential for a query expansion effect is limited, which could explain the nonsignificant difference for the English-French case.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Query Expansion Effect.", "sec_num": "5.7.3" }, { "text": "Unfortunately, we cannot say whether the significant improvement in effectiveness occurs mainly because the probability of giving at least one good translation (which is probably the most important factor for retrieval effectiveness [Kraaij 2002; McNamee and Mayfield 2002] ) is higher for QT or indeed because of the query expansion effect. A simulation experiment is needed to quantify the relative contributions. Still, it is of great practical importance that more (weighted) translations can enhance retrieval effectiveness significantly.", "cite_spans": [ { "start": 233, "end": 246, "text": "[Kraaij 2002;", "ref_id": "BIBREF15" }, { "start": 247, "end": 273, "text": "McNamee and Mayfield 2002]", "ref_id": "BIBREF29" } ], "ref_spans": [], "eq_spans": [], "section": "Query Expansion Effect.", "sec_num": "5.7.3" }, { "text": "A related issue is the question of whether it is more important to prune translations or to weight them. Grefenstette (cf. Section 1) originally pointed out the importance of pruning and weighting translations for dictionary-based CLIR. Pruning was seen as a means of removing unwanted senses in a dictionarybased CLIR application. Our experiments confirm the importance of pruning and weighting, but in a slightly different manner. In a CLIR approach based on a Web translation model, the essential function of pruning is to remove spurious translations. Polluted translation models will result in a very poor retrieval effectiveness. As far as sense disambiguation is concerned, we believe that our CLIR models can handle sense ambiguity quite well. Our best-performing runs, based on the 100K models, have on average seven translations per term! Too much pruning (e.g., best match) is suboptimal. However, the more translation alternatives we add, the more important their relative weighting becomes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pruning and Weighting.", "sec_num": "5.7.4" }, { "text": "We have compared weighted translations (QT) with uniform translation probabilities (QT-EQ). In each of the eight comparisons (four language pairs, two pruning techniques), weighting results in an improved retrieval effectiveness. The difference is significant in six of the eight cases. Differences are not significant for the P < 0.1 English-French and French-English translation models. We think this is due to the small number of translations; a uniform translation probability will not differ radically from the estimated translation probabilities.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pruning and Weighting.", "sec_num": "5.7.4" }, { "text": "The importance of weighting is most evident when the 100K translation models are used. These models yield seven translations on average for each term. The CLIR models based on weighted translations are able to exploit the additional information and show improved effectiveness with respect to the P < 0.1 models. The performance of unweighted CLIR models (QT-EQ and SYN) is seriously impaired by the higher number of translations.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pruning and Weighting.", "sec_num": "5.7.4" }, { "text": "The comparison of the naive dictionary-like replacement method, which does not involve any normalization for the number of translations per term (NAIVE), with QT-EQ shows that normalization (i.e. a minimal probabilistic embedding) is essential. The NAIVE runs have the lowest effectiveness of all variant systems (with significant differences). Interestingly, it seems better to select just the one most probable translation than taking all translations unweighted. 5.7.5 Robustness. We pointed out in the previous section that the weighted models are more robust, in the sense that they can handle a large number of translations. We found, however, that the query model translation method (QT) and the document model translation method (DT) display a considerable difference in robustness to noisy translations. Initially we expected that the DT method (in which the matching takes place in the source language) would yield the best results, since this model has previously proven to be successful for several quite different language pairs (e.g., European languages, Chinese, and Arabic using parallel corpora or dictionaries as translation devices [McNamee and Mayfield 2001; Xu, Weischedel, and Nguyen 2001; Hiemstra et al. 2001] ).", "cite_spans": [ { "start": 1151, "end": 1178, "text": "[McNamee and Mayfield 2001;", "ref_id": "BIBREF28" }, { "start": 1179, "end": 1211, "text": "Xu, Weischedel, and Nguyen 2001;", "ref_id": "BIBREF50" }, { "start": 1212, "end": 1233, "text": "Hiemstra et al. 2001]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Pruning and Weighting.", "sec_num": "5.7.4" }, { "text": "However, our initial DT runs obtained extremely poor results. We discovered that this was largely due to noisy translations from the translation model (pruned by the P < 0.1 or 100K method), which is based on Web data. There are many terms in the target language that occur very rarely in the parallel Web corpus. The translation probabilities for these terms (based on the most probable alignments) are therefore unreliable. Often these rare terms (and nonwords like xc64) are aligned with more common terms in the other language and are not pruned by the default pruning criteria (P > 0.1 or best 100K parameters), since they have high translation probabilities. This especially poses a problem for the DT model, since it includes a summation over all terms in the target language that occur in the document and have a nonzero translation probability. We devised a supplementary pruning criterion to remove these noisy translations, discarding all translations for which the source term has a marginal probability in the translation model that is below a particular value (typically between 10 \u22126 and 10 \u22125 ). Later we discovered that a simple pruning method was even more effective: discarding all translations for which either the source or target term contains a digit. The results in Tables 7 and 8 are based on the latter additional pruning criterion. The QT approach is less sensitive to noisy translations arising from rare terms in the target language, because it is easy to remove these translations using a probability threshold. We deduce that extra care therefore has to be taken to prune translation models for the document model translation approach to CLIR.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Pruning and Weighting.", "sec_num": "5.7.4" }, { "text": "Statistical translation models require large parallel corpora, and unfortunately, only a few manually constructed ones are available. In this article, we have explored the possibility of automatically mining the Web for parallel texts in order to construct such corpora. Translation models are then trained on these corpora. We subsequently examined different ways to embed the resulting translation models in a cross-language information retrieval system.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "To mine parallel Web pages, we constructed a mining system called PTMiner. This system employs a series of heuristics to locate candidate parallel pages and determine whether they are indeed parallel. We have successfully used PTMiner to construct corpora for a number of different language pairs: English-French, English-Italian, English-German, English-Dutch, and English-Chinese. The language-independent characteristics of PTMiner allowed us to adapt it quite easily to different language pairs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "The heuristics used in the mining process seem to be effective. Although the system cannot collect all pairs of parallel pages, our preliminary evaluation shows that its precision is quite high. (The recall ratio is less important in this context because of the abundance of parallel pages on the Web.)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "The mining results-parallel corpora-are subsequently used to train statistical translation models, which are exploited in a CLIR system. The major advantage of this approach is that it can be fully automated, avoiding the tedious work of manual collection of parallel corpora. On the other hand, compared to manually prepared parallel corpora, our mining results contain more noise (i.e., nonparallel pages). For a general translation task this may be problematic; for CLIR, however, the noise contained in the corpora is less dramatic. In fact, IR is strongly error tolerant. A small proportion of incorrect translation words can be admitted without a major impact on global effectiveness. Our experiments showed that a CLIR approach based on the mined Web corpora can in fact outperform a good MT system (Systran). This confirms our initial hypothesis that noisy parallel corpora can be very useful for applications such as CLIR. Our demonstration that the Web can indeed be used as a large parallel corpus for tasks such as CLIR is the main contribution of this article.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "Most previous work on CLIR has separated the translation stage from the retrieval stage (i.e., query translation is considered as a preprocessing step for monolingual IR). In this article, we have integrated translation and retrieval within the same framework. The advantage of this integration is that we do not need to obtain the optimal translation of a source query, and then an optimal retrieval result given a query translation, but instead aim for the optimal global effect. The comparisons between our approach and simulated external approaches clearly show that an integrated approach performs better.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "We also compared two ways of embedding translation models within a CLIR system: (1) translating the source query model into the target (document) language, and (2) translating the document model into the source language. 20 Both embedding methods produced very good results compared to our reference run with Systran. However, it is still too early to assert which embedding method is superior. We did observe a significant difference in robustness between the two methods: The document model translation method is much more sensitive to spurious translations, since the model incorporates into a query term all source terms that have a nonzero translation probability. We devised two supplementary pruning techniques that effectively removed the noisy terms: removing terms containing digits, and removing translations based on source terms with a low marginal probability. (This latter approach is perhaps more principled.)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "On the use of statistical translation models for CLIR, we have demonstrated that this naturally produces a desired query expansion effect, resulting in more related documents being found. In our experimental evaluation, we saw that it is usually better to include more than one translation, and to weigh these translations according to the translation probabilities, rather than using the resulting translation model as a bilingual lexicon for external translation. This effect partly accounts for the success of our approach in comparison with an MT-based approach, which retains only one translation per sense. However, this technique should not be exaggerated; otherwise, too much noise will be introduced. To avoid this, it is important to incorporate pruning.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "We investigated several ways to prune translation models. The best results were obtained with a pruning method based on the top 100K parameters of the translation model. The translation models pruned with the best 100K parameters method produced more than seven translations per word on average, demonstrating the capability of the CLIR model to handle translation ambiguity and exploit co-occurrence information from the parallel corpus for query expansion.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "There are several ways in which our approach can be improved. First, regarding PTMiner, more or better heuristics could be integrated into the mining algorithm. As we mentioned, parallel Web sites are not always organized in the ways we would expect. This is particularly the case for those in non-European languages such as Chinese and Japanese. Hence, one of the questions we wish to investigate is how to extend the coverage of PTMiner to more parallel Web pages. One possible improvement would be to integrate a component that \"learns\" the organization patterns of a particular Web site (assuming, of course, that the Web site is organized in a consistent way). Preliminary tests have shown that this is possible to some extent: We can recognize dynamically that the parallel pages on www.operationid.com are at www.operationcarte.com or that the file index1.html corresponds to index2.html . Such criteria complement the ones currently employed in PTMiner.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "In its current form, PTMiner scans candidates for parallel Web sites according to similarities in their file names. This step does not exploit the hyperlinks between the pages, whereas we know that two pages that are referenced at comparable structural positions in two parallel pages have a very high chance of themselves being parallel. Exploiting hyperlink structure to (help) find parallel Web pages could well improve the quality of PTMiner.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "When the mining results are not fully parallel, it would be interesting to attempt to clean them in order to obtain a higher-quality training material. One possible approach for doing this would be to use sentence alignment as an additional filter, as we mentioned earlier. This approach has been applied successfully to our English-Chinese Web corpus. The cleaned corpus results in both higher translation accuracy and higher CLIR effectiveness. However, this approach has still to be tested for the European languages.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "In this study, we hypothesized that IBM Model 1 is appropriate for CLIR, primarily because word order is not important for IR. Although it is true that word order is not important in current IR approaches, it is definitely important to consider context words during the translation. For example, when deciding how to translate the French word tableau (which may refer to a painting, a blackboard, a table [of data], etc.), if we observe artistique ('artistic') next to it, then it is pretty certain that tableau refers to a painting. A more sophisticated translation model than IBM Model 1 could produce a better selection of translation words.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "We also rely solely on word translation in our approach, although it is well known that this simplistic approach cannot correctly translate compound terms such as pomme de terre ('potato') and cul de sac ('no exit'). Incorporating the translation of compound terms in a translation model should result in additional improvements for CLIR. Our preliminary experiments (Nie and Dufort 2002) on integrating the translation of compounds certainly showed this, with improvement of up to 70% over a word-based approach. This direction warrants further investigation. Finally, all our efforts thus far to mine parallel Web pages have involved English. How can we deal with CLIR between, say, Chinese and German, for which there are few parallel Web sites? One possible solution would be to use English as a pivot language, even though the two-step translation involved would certainly reduce accuracy and introduce more noise. Nevertheless, several authors have shown that a pivot approach can still produce effective retrieval and can at least complement a dictionary-based approach (Franz, McCarley, and Ward 2000; Gollins and Sanderson 2001; Lehtokangas and Airio 2002) .", "cite_spans": [ { "start": 367, "end": 388, "text": "(Nie and Dufort 2002)", "ref_id": "BIBREF34" }, { "start": 1077, "end": 1109, "text": "(Franz, McCarley, and Ward 2000;", "ref_id": null }, { "start": 1110, "end": 1137, "text": "Gollins and Sanderson 2001;", "ref_id": null }, { "start": 1138, "end": 1165, "text": "Lehtokangas and Airio 2002)", "ref_id": "BIBREF26" } ], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6." }, { "text": "We used the free translation service provided at http://babelfish.altavista.com/ in October 2002. 2 Although there is no inherent obstacle preventing MT systems from generating multiple translations, in practice, only one translation is produced.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Precision is defined as the proportion of relevant documents among all the retrieved documents. 4 Recall is the proportion of relevant documents retrieved among all the relevant documents in a collection.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "LDC provides a version containing texts from the mid-1970s through 1988; see http://www.ldc.upenn.edu/ .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The processing on Chinese is described inChen and Nie (2000). 7 Currently distributed by OMSEEK:http://cvs.sourceforge.net/cgi-bin/viewcvs.cgi/omseek/om/languages/ .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The model is referred to as noisy-channel because it takes the view that S is the result of some input signal T's being corrupted while passing through a noisy channel. In this context, the goal is to recover the initial input, given the corrupted output.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The fact that language models have to be smoothed seems to contradict the discussion in Section 3, in which we stated that rare terms are not critical for IR effectiveness, but it actually does not. Smoothing helps to make the distinction between absent important terms (middle-frequency terms) and absent nonimportant terms (high-frequency terms). The score of a document that misses important terms should be lowered more than that of a document that misses an unimportant term.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The NLLR can also be reformulated as a difference of two Kullback-Leibler divergences(Ng 2000). 11 We omit the normalization with the background model in the formula for presentation reasons.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "For presentation reasons, we have replaced the variable \u03c4 used in Section 4.1 with s and t for a term in the source and target language, respectively.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "It may be better to view them as mixing weights in this case.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "A more extreme case is query C044 about the \"tour de france.\" According to the P(e | f ) > 0.1 translation model, there are 902 French words that translate into the \"English\" word de. This is mostly due to French proper names, which are left untranslated in the English parallel text. 15 CLEF = Cross Language Evaluation Forum, www.clef-campaign.org . 16 Topics without relevant documents in a subcollection were discarded. 17 Available at http://www.xrce.xerox.com/competencies/ats/xelda/summary.html .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "We have not been able to substantiate this claim with quantitative figures but did analyze the lemmas that were not found in the translation dictionaries during query translation. We did not find any structural mismatches.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "This figure includes proper nouns.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Another method that interprets multiple translations as synonyms is a special case of the latter.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "This work was partly funded by a research grant from the Dutch Telematics Institute: DRUID project. We would like to thank Xerox Research Center Europe (XRCE) for making its Xelda toolkit available to us. We would also like to thank George Foster for making his statistical MT toolkit available and for many interesting discussions. Special thanks are due to Jiang Chen, who contributed to the building of PTMiner. Finally, we want to thank Elliott Macklovitch and the two anonymous reviewers for their constructive comments and careful review. Part of this work was carried out while the first author was visiting the RALI laboratory at Universit\u00e9 de Montr\u00e9al. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Event-Based Information Organization", "authors": [ { "first": "James", "middle": [], "last": "Allen", "suffix": "" } ], "year": 2002, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Allen, James, editor. 2002. Event-Based Information Organization. Kluwer Academic, Boston.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "An inequality and associated maximization technique in statistical estimations of probabilistic functions of Markov processes", "authors": [ { "first": "L", "middle": [ "E" ], "last": "Baum", "suffix": "" } ], "year": 1972, "venue": "Inequalities", "volume": "3", "issue": "", "pages": "1--8", "other_ids": {}, "num": null, "urls": [], "raw_text": "Baum, L. E. 1972. An inequality and associated maximization technique in statistical estimations of probabilistic functions of Markov processes. Inequalities, 3:1-8.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "National Institute of Standards and Technology Special Publication 500-246", "authors": [ { "first": "Adam", "middle": [], "last": "Berger", "suffix": "" }, { "first": "John", "middle": [], "last": "Lafferty", "suffix": "" } ], "year": 2000, "venue": "Advances in Information Retrieval: 25th European Conference on IR Research (ECIR 2003)", "volume": "8", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Berger, Adam and John Lafferty. 2000. The Weaver system for document retrieval. In Ellen M. Voorhees and Donna K. Harman, editors, The Eighth Text Retrieval Conference (TREC-8), volume 8. National Institute of Standards and Technology Special Publication 500-246, Gaithersburg, MD. Braschler, Martin and B\u00e4rbel Ripplinger. 2003. Stemming and decompounding for German text retrieval. In Fabrizio Sebastiani, editor, Advances in Information Retrieval: 25th European Conference on IR Research (ECIR 2003), Pisa, Italy, April 2003, Proceedings. Lecture Notes in Computer Science 2633. Springer, Berlin.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Document retrieval and routing using the INQUERY system", "authors": [ { "first": "John", "middle": [], "last": "Broglio", "suffix": "" }, { "first": "James", "middle": [ "P" ], "last": "Callan", "suffix": "" }, { "first": "W", "middle": [ "Bruce" ], "last": "Croft", "suffix": "" }, { "first": "Daniel", "middle": [ "W" ], "last": "Nachbar", "suffix": "" } ], "year": 1995, "venue": "The Third Text Retrieval Conference", "volume": "4", "issue": "", "pages": "29--38", "other_ids": {}, "num": null, "urls": [], "raw_text": "Broglio, John, James P. Callan, W. Bruce Croft, and Daniel W. Nachbar. 1995. Document retrieval and routing using the INQUERY system. In Donna K. Harman, editor, The Third Text Retrieval Conference, volume 4. National Institute of Standards and Technology Special Publication 500-236, Gaithersburg, MD, pages 29-38.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "A statistical approach to machine translation", "authors": [ { "first": "Peter", "middle": [ "F" ], "last": "Brown", "suffix": "" }, { "first": "John", "middle": [], "last": "Cocke", "suffix": "" }, { "first": "Stephen", "middle": [ "A Della" ], "last": "Pietra", "suffix": "" }, { "first": "Vincent", "middle": [ "J Della" ], "last": "Pietra", "suffix": "" }, { "first": "Fredrick", "middle": [], "last": "Jelinek", "suffix": "" }, { "first": "John", "middle": [ "D" ], "last": "Lafferty", "suffix": "" }, { "first": "Robert", "middle": [ "L" ], "last": "Mercer", "suffix": "" }, { "first": "Paul", "middle": [ "S" ], "last": "Roossin", "suffix": "" } ], "year": 1990, "venue": "Computational Linguistics", "volume": "16", "issue": "2", "pages": "79--85", "other_ids": {}, "num": null, "urls": [], "raw_text": "Brown, Peter F., John Cocke, Stephen A. Della Pietra, Vincent J. Della Pietra, Fredrick Jelinek, John D. Lafferty, Robert L. Mercer, and Paul S. Roossin. 1990. A statistical approach to machine translation. Computational Linguistics, 16(2):79-85.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Web parallel text mining for Chinese-English cross-language information retrieval", "authors": [ { "first": "Jiang", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Jian-Yun", "middle": [], "last": "Nie", "suffix": "" } ], "year": 2000, "venue": "Proceedings of NAACL-ANLP", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chen, Jiang and Jian-Yun Nie. 2000. Web parallel text mining for Chinese-English cross-language information retrieval. In Proceedings of NAACL-ANLP, Seattle.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "SP-An SGML System Conforming to International Standard ISO 8879-Standard Generalized Markup Language", "authors": [ { "first": "James", "middle": [], "last": "Clark", "suffix": "" } ], "year": 2001, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Clark, James. 2001. SP-An SGML System Conforming to International Standard ISO 8879-Standard Generalized Markup Language. Available at http://www.jclark.com/sp/ .", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Practical Nonparametric Statistics", "authors": [ { "first": "William", "middle": [], "last": "Conover", "suffix": "" }, { "first": "", "middle": [], "last": "Jay", "suffix": "" }, { "first": "London", "middle": [], "last": "Wiley", "suffix": "" }, { "first": "W", "middle": [], "last": "Croft", "suffix": "" }, { "first": "Alistair", "middle": [], "last": "Bruce", "suffix": "" }, { "first": "C", "middle": [ "J" ], "last": "Moffat", "suffix": "" }, { "first": "", "middle": [], "last": "Keith", "suffix": "" } ], "year": 1980, "venue": "Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '98)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Conover, William Jay. 1980. Practical Nonparametric Statistics. Wiley, London. Croft, W. Bruce, Alistair Moffat, C. J. \"Keith\" van Rijsbergen, Ross Wilkinson, and Justin Zobel, editors. 1998. Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '98). ACM Press.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Automatic cross-language retrieval using latent semantic indexing", "authors": [ { "first": "", "middle": [], "last": "Landauer", "suffix": "" } ], "year": 1997, "venue": "AAAI Spring Symposium on on Cross-Language Text and Speech Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Landauer. 1997. Automatic cross-language retrieval using latent semantic indexing. In AAAI Spring Symposium on on Cross-Language Text and Speech Retrieval. American Association for Artificial Intelligence. Available at http://www.aaai.org/Press/Reports/ Symposia/Spring/ss-97-05.html .", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Querying across languages: A dictionary-based approach to multilingual information retrieval", "authors": [ { "first": "David", "middle": [], "last": "Hull", "suffix": "" }, { "first": "Gregory", "middle": [], "last": "Grefenstette", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 19th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '96)", "volume": "", "issue": "", "pages": "49--57", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hull, David and Gregory Grefenstette. 1996. Querying across languages: A dictionary-based approach to multilingual information retrieval. In Hans-Peter Frei, Donna Harman, Peter Sch\u00e4uble, and Ross Wilkinson, editors, Proceedings of the 19th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '96). ACM Press, New York, pages 49-57.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Advanced approaches to the statistical analysis of TREC information retrieval experiments", "authors": [ { "first": "", "middle": [], "last": "Hull", "suffix": "" }, { "first": "Paul", "middle": [ "B" ], "last": "David", "suffix": "" }, { "first": "Kwong Bor", "middle": [], "last": "Kantor", "suffix": "" }, { "first": "", "middle": [], "last": "Ng", "suffix": "" } ], "year": 1999, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hull, David, Paul B. Kantor, and Kwong Bor Ng. 1999. Advanced approaches to the statistical analysis of TREC information retrieval experiments. Unpublished report, Rutgers University, New Brunswick, NJ.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Corpus encoding standard", "authors": [ { "first": "Nancy", "middle": [], "last": "Ide", "suffix": "" }, { "first": "G", "middle": [], "last": "Priest-Dorman", "suffix": "" }, { "first": "Jean", "middle": [], "last": "V\u00e9ronis", "suffix": "" } ], "year": 1995, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ide, Nancy, G. Priest-Dorman and Jean V\u00e9ronis. 1995. Corpus encoding standard. Available at http://www.cs.vassar.edu/ CES/ .", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Searching the Web: The public and their queries", "authors": [ { "first": "Bernard", "middle": [ "J" ], "last": "Jansen", "suffix": "" }, { "first": "Amanda", "middle": [], "last": "Spink", "suffix": "" }, { "first": "Deitmar", "middle": [], "last": "Wolfram", "suffix": "" }, { "first": "Tefko", "middle": [], "last": "Saracevic", "suffix": "" } ], "year": 2001, "venue": "Journal of the American Society for Information Science and Technology", "volume": "53", "issue": "3", "pages": "226--234", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jansen, Bernard J., Amanda Spink, Deitmar Wolfram, and Tefko Saracevic. 2001. Searching the Web: The public and their queries. Journal of the American Society for Information Science and Technology, 53(3):226-234.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "TNO at CLEF-2001: Comparing translation resources", "authors": [ { "first": "Wessel", "middle": [], "last": "Kraaij", "suffix": "" } ], "year": 2001, "venue": "Evaluation of Cross-Language Information Retrieval Systems: Second Workshop of the Cross-Language Evaluation Forum", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kraaij, Wessel. 2002. TNO at CLEF-2001: Comparing translation resources. In Carol Peters, Martin Braschler, Julio Gonzalo, and Michael Kluck, editors, Evaluation of Cross-Language Information Retrieval Systems: Second Workshop of the Cross-Language Evaluation Forum (CLEF 2001). Springer Verlag, Berlin.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Viewing stemming as recall enhancement", "authors": [ { "first": "Wessel", "middle": [], "last": "Kraaij", "suffix": "" }, { "first": "Ren\u00e9e", "middle": [], "last": "Pohlmann", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the 19th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '96)", "volume": "", "issue": "", "pages": "40--48", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kraaij, Wessel and Ren\u00e9e Pohlmann. 1996. Viewing stemming as recall enhancement. In Hans-Peter Frei, Donna Harman, Peter Sch\u00e4uble, and Ross Wilkinson, editors, Proceedings of the 19th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '96). ACM Press, New York, pages 40-48.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Twenty-one at TREC-8: Using language technology for information retrieval", "authors": [ { "first": "Wessel", "middle": [], "last": "Kraaij", "suffix": "" }, { "first": "Ren\u00e9e", "middle": [], "last": "Pohlmann", "suffix": "" }, { "first": "Djoerd", "middle": [], "last": "Hiemstra", "suffix": "" } ], "year": 2000, "venue": "The Eighth Text Retrieval Conference (TREC-8)", "volume": "8", "issue": "", "pages": "500--246", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kraaij, Wessel, Ren\u00e9e Pohlmann, and Djoerd Hiemstra. 2000. Twenty-one at TREC-8: Using language technology for information retrieval. In Ellen M. Voorhees and Donna K. Harman, editors, The Eighth Text Retrieval Conference (TREC-8), volume 8. National Institute of Standards and Technology Special Publication 500-246, Gaithersburg, MD.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Language models for topic tracking", "authors": [ { "first": "Wessel", "middle": [], "last": "Kraaij", "suffix": "" }, { "first": "Martijn", "middle": [], "last": "Spitters", "suffix": "" } ], "year": 2003, "venue": "Language Models for Information Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kraaij, Wessel and Martijn Spitters. 2003. Language models for topic tracking. In Bruce Croft and John Lafferty, editors, Language Models for Information Retrieval. Kluwer Academic, Boston.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "The importance of prior probabilities for entry page search", "authors": [ { "first": "Wessel", "middle": [], "last": "Kraaij", "suffix": "" }, { "first": "Thijs", "middle": [], "last": "Westerveld", "suffix": "" }, { "first": "Djoerd", "middle": [], "last": "Hiemstra", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kraaij, Wessel, Thijs Westerveld, and Djoerd Hiemstra. 2002. The importance of prior probabilities for entry page search. In Micheline Beaulieu, Ricardo Baeza-Yates, Sung Hyon Myaeng, and Kalervo J\u00e4rvelin, editors, Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2002). ACM Press, New York.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "English-Chinese cross-language retrieval based on a translation package", "authors": [ { "first": "K", "middle": [ "L" ], "last": "Kwok", "suffix": "" } ], "year": 1999, "venue": "Workshop: Machine Translation for Cross Language Information Retrieval", "volume": "", "issue": "", "pages": "8--13", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kwok, K. L. 1999. English-Chinese cross-language retrieval based on a translation package. In Workshop: Machine Translation for Cross Language Information Retrieval, Singapore, Machine Translation Summit VII, pages 8-13.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Document language models, query models, and risk minimization for information retrieval", "authors": [ { "first": "John", "middle": [], "last": "Lafferty", "suffix": "" }, { "first": "Chengxiang", "middle": [], "last": "Zhai", "suffix": "" } ], "year": 2001, "venue": "Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lafferty, John and Chengxiang Zhai. 2001a. Document language models, query models, and risk minimization for information retrieval. In W. Bruce Croft, David J. Harper, Donald H. Kraft, and Justin Zobel, editors, Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2001). ACM Press, New York.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Probabilistic IR models based on document and query generation", "authors": [ { "first": "John", "middle": [], "last": "Lafferty", "suffix": "" }, { "first": "Chengxiang", "middle": [], "last": "Zhai", "suffix": "" } ], "year": 2001, "venue": "Proceedings of the Workshop on Language Modeling and Information Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lafferty, John and Chengxiang Zhai. 2001b. Probabilistic IR models based on document and query generation. In Jamie Callan, Bruce Croft, and John Lafferty, editors, Proceedings of the Workshop on Language Modeling and Information Retrieval, Pittsburgh.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "On constructing a transfer dictionary for man and machine", "authors": [ { "first": "John", "middle": [], "last": "Laffling", "suffix": "" } ], "year": 1992, "venue": "Target", "volume": "4", "issue": "1", "pages": "17--31", "other_ids": {}, "num": null, "urls": [], "raw_text": "Laffling, John. 1992. On constructing a transfer dictionary for man and machine. Target, 4(1):17-31.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Cross-lingual relevance models", "authors": [ { "first": "Victor", "middle": [], "last": "Lavrenko", "suffix": "" }, { "first": "Martin", "middle": [], "last": "Choquette", "suffix": "" }, { "first": "W", "middle": [ "Bruce" ], "last": "Croft", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lavrenko, Victor, Martin Choquette, and W. Bruce Croft. 2002. Cross-lingual relevance models. In Micheline Beaulieu, Ricardo Baeza-Yates, Sung Hyon Myaeng, and Kalervo J\u00e4rvelin, editors, Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2002). ACM Press, New York.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "Relevance-based language models", "authors": [ { "first": "Victor", "middle": [], "last": "Lavrenko", "suffix": "" }, { "first": "W. Bruce", "middle": [], "last": "Croft", "suffix": "" } ], "year": 2001, "venue": "Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lavrenko, Victor and W. Bruce Croft. 2001. Relevance-based language models. In W. Bruce Croft, David J. Harper, Donald H. Kraft, and Justin Zobel, editors, Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2001). ACM Press, New York.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "Translation via a pivot language challenges direct translation in CLIR", "authors": [ { "first": "Raija", "middle": [], "last": "Lehtokangas", "suffix": "" }, { "first": "Eija", "middle": [], "last": "Airio", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the SIGIR 2002 Workshop: Cross-Language Information Retrieval: A Research Roadmap", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lehtokangas, Raija and Eija Airio. 2002. Translation via a pivot language challenges direct translation in CLIR. In Proceedings of the SIGIR 2002 Workshop: Cross-Language Information Retrieval: A Research Roadmap, Tampere, Finland.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "Parallel text collections at the Linguistic Data Consortium", "authors": [ { "first": "Xiaoyi", "middle": [], "last": "Ma", "suffix": "" } ], "year": 1999, "venue": "Machine Translation Summit VII", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ma, Xiaoyi. 1999. Parallel text collections at the Linguistic Data Consortium. In Machine Translation Summit VII, Singapore.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "Cross-Language Information Retrieval and Evaluation. Lecture Notes in Computer Science 2069", "authors": [ { "first": "Paul", "middle": [], "last": "Mcnamee", "suffix": "" }, { "first": "James", "middle": [], "last": "Mayfield", "suffix": "" } ], "year": 2001, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "McNamee, Paul and James Mayfield. 2001. A language-independent approach to European text retrieval. In Carol Peters, editor, Cross-Language Information Retrieval and Evaluation. Lecture Notes in Computer Science 2069. Springer Verlag, Berlin.", "links": null }, "BIBREF29": { "ref_id": "b29", "title": "Comparing cross-language query expansion techniques by degrading translation reources", "authors": [ { "first": "Paul", "middle": [], "last": "Mcnamee", "suffix": "" }, { "first": "James", "middle": [], "last": "Mayfield", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "McNamee, Paul and James Mayfield. 2002. Comparing cross-language query expansion techniques by degrading translation reources. In Micheline Beaulieu, Ricardo Baeza-Yates, Sung Hyon Myaeng, and Kalervo J\u00e4rvelin, editors, Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2002). ACM Press, New York.", "links": null }, "BIBREF30": { "ref_id": "b30", "title": "A hidden Markov model information retrieval system", "authors": [ { "first": "David", "middle": [ "R H" ], "last": "Miller", "suffix": "" }, { "first": "Tim", "middle": [], "last": "Leek", "suffix": "" }, { "first": "Richard", "middle": [ "M" ], "last": "Schwartz", "suffix": "" } ], "year": 1999, "venue": "Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '99)", "volume": "", "issue": "", "pages": "214--221", "other_ids": {}, "num": null, "urls": [], "raw_text": "Miller, David R. H., Tim Leek, and Richard M. Schwartz. 1999. A hidden Markov model information retrieval system. In Marti Hearst, Fred Gey, and Richard Tong, editors, Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '99). ACM Press, New York, pages 214-221.", "links": null }, "BIBREF31": { "ref_id": "b31", "title": "A maximum likelihood ratio information retrieval model", "authors": [ { "first": "Kenney", "middle": [], "last": "Ng", "suffix": "" } ], "year": 2000, "venue": "The Eighth Text Retrieval Conference", "volume": "8", "issue": "", "pages": "500--246", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ng, Kenney. 2000. A maximum likelihood ratio information retrieval model. In Ellen M. Voorhees and Donna K. Harman, editors, The Eighth Text Retrieval Conference (TREC-8), volume 8. National Institute of Standards and Technology Special Publication 500-246, Gaithersburg, MD.", "links": null }, "BIBREF32": { "ref_id": "b32", "title": "Query expansion and query translation as logical inference", "authors": [ { "first": "Jian-Yun", "middle": [], "last": "Nie", "suffix": "" } ], "year": 2002, "venue": "Journal of the American Society for Information Science and Technology", "volume": "54", "issue": "4", "pages": "340--351", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nie, Jian-Yun. 2002. Query expansion and query translation as logical inference. Journal of the American Society for Information Science and Technology, 54(4): 340-351.", "links": null }, "BIBREF33": { "ref_id": "b33", "title": "Filtering noisy parallel corpora of Web pages", "authors": [ { "first": "Jian-Yun", "middle": [], "last": "Nie", "suffix": "" }, { "first": "Jian", "middle": [], "last": "Cai", "suffix": "" } ], "year": 2001, "venue": "IEEE Symposium on NLP and Knowledge Engineering", "volume": "", "issue": "", "pages": "453--458", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nie, Jian-Yun and Jian Cai. 2001. Filtering noisy parallel corpora of Web pages. In IEEE Symposium on NLP and Knowledge Engineering, Tucson, AZ, pages 453-458.", "links": null }, "BIBREF34": { "ref_id": "b34", "title": "Combining words and compound terms for monolingual and cross-language information retrieval", "authors": [ { "first": "Jian-Yun", "middle": [], "last": "Nie", "suffix": "" }, { "first": "Jean-Fran\u00e7ois", "middle": [], "last": "Dufort", "suffix": "" } ], "year": 2002, "venue": "Proceedings of Information", "volume": "", "issue": "", "pages": "453--458", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nie, Jian-Yun and Jean-Fran\u00e7ois Dufort. 2002. Combining words and compound terms for monolingual and cross-language information retrieval. In Proceedings of Information 2002, Beijing, pages 453-458.", "links": null }, "BIBREF35": { "ref_id": "b35", "title": "Cross-language information retrieval based on parallel texts and automatic mining of parallel texts from the Web", "authors": [ { "first": "Jian-Yun", "middle": [], "last": "Nie", "suffix": "" }, { "first": "Michel", "middle": [], "last": "Simard", "suffix": "" }, { "first": "Pierre", "middle": [], "last": "Isabelle", "suffix": "" }, { "first": "Richard", "middle": [], "last": "Durand", "suffix": "" } ], "year": 1999, "venue": "Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '99)", "volume": "", "issue": "", "pages": "74--81", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nie, Jian-Yun, Michel Simard, Pierre Isabelle, and Richard Durand. 1999. Cross-language information retrieval based on parallel texts and automatic mining of parallel texts from the Web. In Marti Hearst, Fred Gey, and Richard Tong, editors, Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '99). ACM Press, New York, pages 74-81.", "links": null }, "BIBREF36": { "ref_id": "b36", "title": "The effects of query structure and dictionary setups in dictionary-based cross-language information retrieval", "authors": [ { "first": "Ari", "middle": [], "last": "Pirkola", "suffix": "" } ], "year": 1998, "venue": "Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '98)", "volume": "", "issue": "", "pages": "55--63", "other_ids": {}, "num": null, "urls": [], "raw_text": "Pirkola, Ari. 1998. The effects of query structure and dictionary setups in dictionary-based cross-language information retrieval. In W. Bruce Croft, Alistair Moffat, C. J. \"Keith\" van Rijsbergen, Ross Wilkinson, and Justin Zobel, editors, Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '98). ACM Press, New York, pages 55-63.", "links": null }, "BIBREF37": { "ref_id": "b37", "title": "A language modeling approach to information retrieval", "authors": [ { "first": "Jay", "middle": [ "M" ], "last": "Ponte", "suffix": "" }, { "first": "W", "middle": [ "Bruce" ], "last": "Croft", "suffix": "" } ], "year": 1998, "venue": "Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '98)", "volume": "", "issue": "", "pages": "275--281", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ponte, Jay M. and W. Bruce Croft. 1998. A language modeling approach to information retrieval. In W. Bruce Croft, Alistair Moffat, C. J. \"Keith\" van Rijsbergen, Ross Wilkinson, and Justin Zobel, editors, Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '98). ACM Press, New York, pages 275-281.", "links": null }, "BIBREF38": { "ref_id": "b38", "title": "An algorithm for suffix stripping", "authors": [ { "first": "Martin", "middle": [ "F" ], "last": "Porter", "suffix": "" } ], "year": 1980, "venue": "", "volume": "14", "issue": "", "pages": "130--137", "other_ids": {}, "num": null, "urls": [], "raw_text": "Porter, Martin F. 1980. An algorithm for suffix stripping. Program, 14(3):130-137.", "links": null }, "BIBREF39": { "ref_id": "b39", "title": "Clean up your Web pages with HTML TIDY", "authors": [ { "first": "Dave", "middle": [], "last": "Ragget", "suffix": "" } ], "year": 1998, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ragget, Dave. 1998. Clean up your Web pages with HTML TIDY. Available at http://www.w3.org/People/ Raggett/tidy/ .", "links": null }, "BIBREF40": { "ref_id": "b40", "title": "Parallel stands: A preliminary investigation into mining the Web for bilingual text", "authors": [ { "first": "Philip", "middle": [], "last": "Resnik", "suffix": "" } ], "year": 1998, "venue": "Proceedings of AMTA. Lecture Notes in Computer Science 1529", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Resnik, Philip. 1998. Parallel stands: A preliminary investigation into mining the Web for bilingual text. In Proceedings of AMTA. Lecture Notes in Computer Science 1529. Springer, Berlin.", "links": null }, "BIBREF41": { "ref_id": "b41", "title": "Some simple effective approximations to the 2-Poisson model for probabilistic weighted retrieval", "authors": [ { "first": "Stephen", "middle": [ "E" ], "last": "Robertson", "suffix": "" }, { "first": "Steve", "middle": [], "last": "Walker", "suffix": "" } ], "year": 1994, "venue": "Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '94)", "volume": "", "issue": "", "pages": "232--241", "other_ids": {}, "num": null, "urls": [], "raw_text": "Robertson, Stephen E. and Steve Walker. 1994. Some simple effective approximations to the 2-Poisson model for probabilistic weighted retrieval. In W. Bruce Croft and C. J. \"Keith\" van Rijsbergen, editors, Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '94), pages 232-241. ACM Press, New York.", "links": null }, "BIBREF42": { "ref_id": "b42", "title": "Introduction to Modern Information Retrieval", "authors": [ { "first": "G", "middle": [], "last": "Salton", "suffix": "" }, { "first": "M", "middle": [ "J" ], "last": "Mcgill", "suffix": "" } ], "year": 1983, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Salton, G. and M. J. McGill. 1983. Introduction to Modern Information Retrieval. McGraw-Hill, New York.", "links": null }, "BIBREF43": { "ref_id": "b43", "title": "Report on CLEF-2001 experiments", "authors": [ { "first": "Jacques", "middle": [], "last": "Savoy", "suffix": "" } ], "year": 2001, "venue": "Evaluation of Cross-Language Information Retrieval Systems: Second Workshop of the Cross-Language Evaluation Forum", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Savoy, Jacques. 2002. Report on CLEF-2001 experiments. In Carol Peters, Martin Braschler, Julio Gonzalo, and Michael Kluck, editors, Evaluation of Cross-Language Information Retrieval Systems: Second Workshop of the Cross-Language Evaluation Forum (CLEF 2001). Springer Verlag, Berlin.", "links": null }, "BIBREF44": { "ref_id": "b44", "title": "Building a large multilingual text collection from comparable news documents", "authors": [ { "first": "Paraic", "middle": [], "last": "Sheridan", "suffix": "" }, { "first": "Jean", "middle": [ "Paul" ], "last": "Ballerini", "suffix": "" }, { "first": "Peter", "middle": [], "last": "Sch\u00e4uble", "suffix": "" } ], "year": 1998, "venue": "Cross-Language Information Retrieval. Kluwer Academic", "volume": "", "issue": "", "pages": "137--150", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sheridan, Paraic, Jean Paul Ballerini, and Peter Sch\u00e4uble. 1998. Building a large multilingual text collection from comparable news documents. In Gregory Grefenstette, editor, Cross-Language Information Retrieval. Kluwer Academic, pages 137-150.", "links": null }, "BIBREF45": { "ref_id": "b45", "title": "Using Cognates to Align Sentences in Bilingual Corpora", "authors": [ { "first": "Michel", "middle": [], "last": "Simard", "suffix": "" }, { "first": "George", "middle": [], "last": "Foster", "suffix": "" }, { "first": "Pierre", "middle": [], "last": "Isabelle", "suffix": "" } ], "year": 1992, "venue": "Proceedings of the Fourth Conference on Theoretical and Methodological Issues in Machine Translation (TMI)", "volume": "", "issue": "", "pages": "67--82", "other_ids": {}, "num": null, "urls": [], "raw_text": "Simard, Michel, George Foster, and Pierre Isabelle. 1992. Using Cognates to Align Sentences in Bilingual Corpora. In Proceedings of the Fourth Conference on Theoretical and Methodological Issues in Machine Translation (TMI), pages 67-82, Montr\u00e9al, Qu\u00e9bec.", "links": null }, "BIBREF46": { "ref_id": "b46", "title": "Using language models for tracking events of interest over time", "authors": [ { "first": "Martijn", "middle": [], "last": "Spitters", "suffix": "" }, { "first": "Wessel", "middle": [], "last": "Kraaij", "suffix": "" } ], "year": 2001, "venue": "Proceedings of the Workshop on Language Models for Information Retrieval (LMIR2001)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Spitters, Martijn and Wessel Kraaij. 2001. Using language models for tracking events of interest over time. In Proceedings of the Workshop on Language Models for Information Retrieval (LMIR2001), Pittsburgh.", "links": null }, "BIBREF47": { "ref_id": "b47", "title": "A statistical analysis of the TREC-3 data", "authors": [ { "first": "Jean", "middle": [], "last": "Tague-Sutcliffe", "suffix": "" }, { "first": "James", "middle": [], "last": "Blustein", "suffix": "" } ], "year": 1995, "venue": "The Third Text Retrieval Conference", "volume": "4", "issue": "", "pages": "385--398", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tague-Sutcliffe, Jean and James Blustein. 1995. A statistical analysis of the TREC-3 data. In Donna K. Harman, editor, The Third Text Retrieval Conference, volume 4. National Institute of Standards and Technology Special Publication 500-236, Gaithersburg, MD, pages 385-398.", "links": null }, "BIBREF49": { "ref_id": "b49", "title": "Variations in relevance judgements and the measurement of retrieval effectiveness", "authors": [ { "first": "Ellen", "middle": [ "M" ], "last": "Voorhees", "suffix": "" } ], "year": 1998, "venue": "Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '98)", "volume": "", "issue": "", "pages": "315--323", "other_ids": {}, "num": null, "urls": [], "raw_text": "Voorhees, Ellen M. 1998. Variations in relevance judgements and the measurement of retrieval effectiveness. In W. Bruce Croft, Alistair Moffat, C. J. \"Keith\" van Rijsbergen, Ross Wilkinson, and Justin Zobel, editors, Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '98). ACM Press, New York. pages 315-323.", "links": null }, "BIBREF50": { "ref_id": "b50", "title": "Evaluating a probabilistic model for cross-lingual information retrieval", "authors": [ { "first": "Jinxi", "middle": [], "last": "Xu", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Weischedel", "suffix": "" }, { "first": "Chanh", "middle": [], "last": "Nguyen", "suffix": "" } ], "year": 2001, "venue": "Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Xu, Jinxi, Ralph Weischedel, and Chanh Nguyen. 2001. Evaluating a probabilistic model for cross-lingual information retrieval. In W. Bruce Croft, David J. Harper, Donald H. Kraft, and Justin Zobel, editors, Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2001). ACM Press, New York.", "links": null }, "BIBREF51": { "ref_id": "b51", "title": "Translingual information retrieval: Learning from bilingual corpora", "authors": [ { "first": "Yiming", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Jaime", "middle": [ "G" ], "last": "Carbonell", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Brown", "suffix": "" }, { "first": "Robert", "middle": [ "E" ], "last": "Frederking", "suffix": "" } ], "year": 1998, "venue": "Artificial Intelligence Journal", "volume": "103", "issue": "1-2", "pages": "323--345", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yang, Yiming, Jaime G. Carbonell, Ralph Brown, and Robert E. Frederking. 1998. Translingual information retrieval: Learning from bilingual corpora. Artificial Intelligence Journal, 103(1-2):323-345.", "links": null }, "BIBREF52": { "ref_id": "b52", "title": "Two-stage language models for information retrieval", "authors": [ { "first": "Chengxiang", "middle": [], "last": "Zhai", "suffix": "" }, { "first": "John", "middle": [], "last": "Lafferty", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhai, ChengXiang and John Lafferty. 2002. Two-stage language models for information retrieval. In Micheline Beaulieu, Ricardo Baeza-Yates, Sung Hyon Myaeng, and Kalervo J\u00e4rvelin, editors, Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2002). ACM Press, New York.", "links": null } }, "ref_entries": { "TABREF0": { "num": null, "content": "<table><tr><td/><td colspan=\"3\">English-French English-German English-Italian</td></tr><tr><td>Number of pairs</td><td>18,807</td><td>10,200</td><td>8,504</td></tr><tr><td>Size (MB)</td><td>174/198</td><td>77/100</td><td>50/68</td></tr><tr><td colspan=\"2\">Number of words (M) 6.7/7.1</td><td>1.8/1.8</td><td>1.2/1.3</td></tr><tr><td/><td/><td>English-Dutch</td><td>English-Chinese</td></tr><tr><td/><td/><td>24,738</td><td>14,820</td></tr><tr><td/><td/><td>n.a.</td><td>74/51</td></tr><tr><td/><td/><td>n.a.</td><td>9.2/9.9</td></tr></table>", "text": "Automatically mined corpora. n.a. = not available.", "type_str": "table", "html": null }, "TABREF1": { "num": null, "content": "<table><tr><td/><td colspan=\"2\">English-French English-Italian</td></tr><tr><td>Number of 1-1 alignments</td><td>1018K</td><td>196K</td></tr><tr><td>Number of tokens</td><td>6.7M/7.1M</td><td>1.2M/1.3M</td></tr><tr><td>Number of unique stems</td><td>200K/173K</td><td>102K/87K</td></tr></table>", "text": "Sentence-aligned corpora.", "type_str": "table", "html": null }, "TABREF2": { "num": null, "content": "<table><tr><td>Symbol</td><td>Explanation</td></tr><tr><td>Q</td><td>Query has representation Q = {T 1 , T 2 , . . . , Tn}</td></tr><tr><td>D</td><td>Query has representation D = {T 1 , T 2 , . . . , Tn}</td></tr><tr><td>M Q</td><td>Query language model</td></tr><tr><td>M D</td><td>Document language model</td></tr><tr><td>M C</td><td>Background language model</td></tr><tr><td>\u03c4 i</td><td>index term</td></tr><tr><td>s i</td><td>term in the source language</td></tr><tr><td>t i</td><td>term in the target language</td></tr><tr><td>\u03bb</td><td>smoothing parameter</td></tr><tr><td>c(x)</td><td/></tr></table>", "text": "Common symbols used in equations (1)-(12) and their explanations.", "type_str": "table", "html": null }, "TABREF3": { "num": null, "content": "<table><tr><td>Matching</td></tr></table>", "text": "Explanation of run tags.", "type_str": "table", "html": null }, "TABREF4": { "num": null, "content": "<table><tr><td/><td>French</td><td>English</td><td>Italian</td></tr><tr><td>Document Source</td><td>Le Monde</td><td>Los Angeles Times</td><td>La Stampa</td></tr><tr><td>Number of documents</td><td>44,013</td><td>110,250</td><td>58,051</td></tr><tr><td>Number of topics</td><td>124</td><td>122</td><td>125</td></tr><tr><td>Number of relevant documents</td><td>1,189</td><td>2,256</td><td>1,878</td></tr></table>", "text": "Statistics on the test collection.", "type_str": "table", "html": null }, "TABREF5": { "num": null, "content": "<table><tr><td/><td>English-</td><td>French-</td><td>English-</td><td>Italian-</td></tr><tr><td>Run ID</td><td>French</td><td>English</td><td>Italian</td><td>English</td></tr><tr><td>MONO</td><td>0.4233</td><td>0.4705</td><td>0.4542</td><td>0.4705</td></tr><tr><td>MT</td><td>0.3478</td><td>0.4043</td><td>0.3060</td><td>0.3249</td></tr><tr><td>QT</td><td>a:0.3760</td><td>a:0.4126</td><td>a,b:0.3298</td><td>a:0.3526</td></tr><tr><td>DT</td><td>a:0.3677</td><td>a,b:0.4090</td><td>a:0.3386</td><td>a,b:0.3328</td></tr><tr><td>SYN</td><td>a:0.3730</td><td>b,c:0.3987</td><td>a,b:0.3114</td><td>b:0.3498</td></tr><tr><td>QT-EQ</td><td>a:0.3554</td><td>a,b:0.3987</td><td>c,d:0.3035</td><td>b,c:0.3299</td></tr><tr><td>QT-BM</td><td>a:0.3463</td><td>c,d:0.3769</td><td>b,c:0.3213</td><td>b:0.3221</td></tr><tr><td>NAIVE</td><td>b:0.3303</td><td>d:0.3596</td><td>d:0.2881</td><td>c:0.3183</td></tr><tr><td>Percentage of missed forward</td><td>9.6</td><td>13.54</td><td>16.79</td><td>9.17</td></tr><tr><td>Percentage of missed reverse</td><td>9.08</td><td>14.04</td><td>15.48</td><td>11.31</td></tr><tr><td>Number of translations forward</td><td>1.65</td><td>1.66</td><td>1.86</td><td>2.13</td></tr><tr><td>Number of translations reverse</td><td>22.72</td><td>29.6</td><td>12.00</td><td>22.95</td></tr><tr><td>Table 8</td><td/><td/><td/><td/></tr><tr><td colspan=\"4\">Mean average precision and translation statistics (best 100K parameters).</td><td/></tr><tr><td/><td>English-</td><td>French-</td><td>English-</td><td>Italian-</td></tr><tr><td>Run ID</td><td>French</td><td>English</td><td>Italian</td><td>English</td></tr><tr><td>MONO</td><td>0.4233</td><td>0.4705</td><td>0.4542</td><td>0.4705</td></tr><tr><td>MT</td><td>0.3478</td><td>0.4043</td><td>0.3060</td><td>0.3249</td></tr><tr><td>DT</td><td>a:0.3909</td><td>a:0.4073</td><td>a:0.3728</td><td>a:0.3547</td></tr><tr><td>QT</td><td>a,b:0.3878</td><td>a:0.4194</td><td>a:0.3519</td><td>a:0.3678</td></tr><tr><td>QT-BM</td><td>b:0.3436</td><td>b:0.3702</td><td>b:0.3236</td><td>b:0.3124</td></tr><tr><td>SYN</td><td>c:0.3270</td><td>b:0.3643</td><td>b:0.2958</td><td>c:0.2808</td></tr><tr><td>QT-EQ</td><td>c:0.3102</td><td>b:0.3725</td><td>c:0.2602</td><td>c:0.2595</td></tr><tr><td>NAIVE</td><td>d:0.2257</td><td>c:0.2329</td><td>d:0.2281</td><td>d:0.2021</td></tr><tr><td>Percentage of missed forward</td><td>11.04</td><td>14.65</td><td>16.06</td><td>9.36</td></tr><tr><td>Percentage of missed reverse</td><td>10.39</td><td>16.81</td><td>15.76</td><td>10.53</td></tr><tr><td>Number of translations forward</td><td>7.04</td><td>7.00</td><td>6.36</td><td>7.23</td></tr><tr><td>Number of translations reverse</td><td>10.51</td><td>12.34</td><td>13.32</td><td>17.20</td></tr></table>", "text": "Mean average precision and translation statistics (p > 0.1).", "type_str": "table", "html": null }, "TABREF6": { "num": null, "content": "<table><tr><td>Run ID</td><td>English-French</td><td>French-English</td><td>English-Italian</td><td>Italian-English</td></tr><tr><td>MONO</td><td>0.4233</td><td>0.4705</td><td>0.4542</td><td>0.4705</td></tr><tr><td>MT</td><td>0.3478 (82%)</td><td>0.4043 (86%)</td><td>0.3060 (67%)</td><td>0.3249 (69%)</td></tr><tr><td>DT+QT</td><td>0.4042 (96%)</td><td>0.4273 (87%)</td><td>0.3837 (84%)</td><td>0.3785 (80%)</td></tr></table>", "text": "Mean average precision of combination run, compared to baselines.", "type_str": "table", "html": null } } } }